• The Big Read: Investor Anjney Midha cut Anthropic an early check. His hot streak has kept going • Plus, Recommendations—our weekly pop culture picks: “ Our Town ,” “ Profits, Prophets, Coaches, and Kings ” and “ The Gentlemen ” A hotel room’s quality can be measured in terms of abundance: extra towels, spare bathrobes—plenty of Nespresso pods. Multiples of everything are almost always better. But what’s less pleasant is to inadvertently end up with multiple hotel rooms . A couple days ago, though, I found myself with the unfortunate opportunity to occupy a pair of accommodations at a Santa Monica, Calif., Marriott thanks to an error by my lil’ digital buddy: Muse, Meta Platforms’ new personal agent. The cumbersome, vexing tool hopes to seize on Silicon Valley’s zeitgeisty enthusiasm for popularizing autonomous AI. After a few days with it, I can’t imagine it’s the technology that will take agents mainstream. That hotel snafu came when I gave Muse the hotel’s name and asked it to book two nights. I plugged in my credit card details—a circuitous process that involved both Chase and Stripe—and was met with the digital equivalent of a shrug: An error message informed me that Muse hadn’t completed the transaction, and while it really didn’t know why it couldn’t, it assured me that my credit card hadn’t been charged. A little human skepticism led me to double-check its promise—good thing I did. In fact, the charge had gone through. What happened next still intrigues me. I gave Muse a screenshot showing the charge on my credit card and told it that the charge had gone through. Lickety-split, it returned with a Marriott confirmation number. Not bad—it finished the task without being instructed to do so. But somehow the AI managed to make two bookings: When I showed up to the hotel, I found two rooms waiting for me. The front-desk clerk showed some mercy and canceled one. Had a little human kindness not prevailed, I would’ve been out an extra $408, plus taxes and fees. I’d been pretty gung-ho to try Muse. We’ve spent the past year in increasingly complex discussions about agents ever since OpenClaw came out in January , marking what has felt like the beginning of a distinct new chapter in the AI era. Most of the agents that initially captured attention were meant to simplify work tasks. More recently, a couple of startups have captured buzz with ones devoted to improving life outside the office: The most talked-about one is from Instinct, a startup founded last year. (Instinct is so popular that it now faces a profound compute shortage, and it has recently been seeking additional fundraising that would value it at around $10 billion. Just a month ago, it was valued at $2.5 billion.) But Instinct isn’t widely available, so Muse felt like my first real chance to see what a personal agent could do. The recent reporting from my colleague Jyoti Mann made me even more curious to play around with it: Clearly, Meta sees Muse as a major new product , one aimed at the same everyday people who log on to Instagram and Facebook by the billions. Still, if their experiences are anything like mine, they’ll find Muse something of a misery. The hotel problem wasn’t the only hiccup I ran into with Muse. When I initially set up a account while on my work laptop using my phone number, I later couldn’t access it on my Mac mini at home or my iPhone: Login codes sent via text led me nowhere—just to more error messages. (Frankly, it’s a marvel I found the Muse app at all earlier this week, buried as it was below several other apps also named Muse in Apple’s App Store; it has since risen to the top.) Eventually, I caved and created a second account—this one linked to my Facebook account, which itself is tied to my phone number. Why couldn’t Muse pick up on the overlapping connections? I wish it could have. Later, I did get Muse to complete a Resy reservation and schedule an Uber. I can’t truthfully tell you it was faster or easier than if I’d just gone directly to those apps. After a while, I hit an existential conundrum. I’d given it the tasks that occurred to me, which numbered a small handful. What else could I get Muse to do? I’ve found interacting with the AI something like trying to wrangle a lackluster employee. If I wanted to maximize its potential, I’d need to think deeply and creatively about what else it could possibly do, coax it and baby it, then bite my nails and hope it actually carried out what I wanted. As I relate these frustrations, I can already hear someone shouting “Skill issue!” at me. But really, I stand by the conviction that mass-market consumer technology shouldn’t require any technical savvy or a lot of effort. Certainly, the best versions of such products do not—even the early versions. If three cars had routinely shown up each time someone ordered their very first Uber ride back in the 2010s, I promise you we’d see more yellow taxis on the road today and fewer Ubers. I don’t see agentic AI as some passing fad, nor am I hoping for such an outcome. The technology’s promise—to automate away some of life’s tedium—is damn alluring. What I expect will happen is that agentic AI will get woven into many existing apps, just as chatbots already populate the internet. People will make regular use of agentic software without ever really knowing it; Anthropic’s Claude, OpenAI’s ChatGPT and Google’s Gemini all already have some agentic capabilities. Obviously, Apple and Google will want to use the technology to make iPhone and Android phones smarter and more useful, and the agents will sync up with what those devices already know and store about us. That would reduce quite a few hurdles to agentic AI. But will new stand-alone apps like Muse take off? I have my doubts—not unless they get much, much simpler and more reliable. Or maybe I’ll just need to learn to see the silver lining in their mistakes. You know what? If I ever again find myself billed twice and double-booked by accident, I know just what to do: Throw a rager in one room, sleep until noon in the other. Maybe Muse can handle ordering the booze. A good party can never have too much. —Abram Brown ( [email protected] ) Weekend’s Latest Stories The Big Read Early Anthropic Investor Seeks VC Glory With Cash and Compute Anjney Midha, 34, wants to get chips in the hands of fledgling startups and academics while he bets on a moment in tech that he describes as the “revenge of the scientists.” Listening: “ Our Town ” The good folks of Gainesboro, Tenn. (population: around 900), know that plenty of outsiders look down their noses at their part of the world—figuring them to be a buncha “backwards, hateful racists,” as one longtime Gainesboro resident puts it. That impression of the place does certainly seem to be why Gainesboro found itself under siege a few years ago when a group of wealthy Christian nationalists started to buy up property in secret, hoping to transform the town into a hotbed for other Christian nationalists: They figured they’d encounter little pushback, especially considering how far their money could go in such an impoverished parish. (I could describe them as white supremacists, but the Christian nationalists do take great umbrage with that label.) “Our Town,” a fast-paced and confidently told podcast from Bloomberg and iHeartRadio, looks at how those Christian nationalists staged their attempted takeover of Gainesboro and how they accumulated their wealth and followers, the latter largely through podcasts. (Sigh—the internet.) It also documents the swift, spirited resistance Gainesboro mounted against them, which cheers the soul and asks us to revisit at least a few of the assumptions about America—and our fellow Americans—that may have become entrenched in our minds. —Abram Brown Reading: “ Profits, Prophets, Coaches, and Kings ” by Jared Diamond Many billions of dollars and quite a few Harvard Business School classes have gone toward trying to figure out what makes a great leader—and turning oafs into passable leaders. Now Jared Diamond, author of the Pulitzer-winning 1998 bestseller “Guns, Germs and Steel,” has set himself the task of defining the essential qualities behind leadership in his latest book, “Profits, Prophets, Coaches, and Kings.” He does so by examining famous figures in business, religion, sports and politics, marking what is surely not the only instance in which Elon Musk has been compared to Genghis Khan. In terms of capitalistic chieftains, Diamond finds that the most distinctive leaders are the ones who’ve had the benefit of both exquisite timing and ruthless execution: Jeff Bezos, for example, outmuscling his competitors in the internet’s Paleozoic Era. As far as politicians go, Diamond points out that the ones we truly remember most aren’t those that simply took a mandate from voters and carried it out, even though we so often say that’s exactly what we want our elected officials to do. Rather, they have championed their own bold ideas and convinced the plebs it was really all part of what they’d originally wanted. All of these conclusions are delivered with Diamond’s dry-humored wonkishness. I enjoyed one in particular: “Once one has decided that one is uniquely qualified to assume the burden of leadership,” Diamond writes, “one’s opinion of oneself is unlikely to change.” —A.B. Watching: “ The Gentlemen ” One of the great joys of “The Gentlemen,” the rollicking “Downton Abbey” meets “The Godfather” concoction from director Guy Ritchie, is to admire how nice Theo James looks as he struts around in magnificent tweed. James plays Eddie Horniman, the fictional Duke of Halstead, and as the Netflix series’ second season begins, it’s just as well that Eddie’s concentrating hard on restoring the family fortune to its fullest extent. The dry-cleaning bills must be enormous: His beautiful wool keeps getting splattered in blood. The first season of “The Gentlemen” found Eddie settling somewhat uncomfortably into his new role: paterfamilias of an old, down-on-its-luck aristocratic family that—much to his surprise—has tied its fortunes to an illicit marijuana operation. In this latest season, Eddie has found his footing and is thinking expansively, which puts him at odds with his business partners: a family of Cockney-accented gangsters—father Bobby (Ray Winstone), who operates from lightly monitored house arrest, and his two children, the very capable Susie (Kaya Scodelario) and strapping, bruised-knuckle Jack (Harry Goodwins). As Eddie’s ambitions get bigger, the stakes get higher and the mood grows darker. And since this is very much a Ritchie production, schemes and mayhem continue to fill the screen in manic fashion. One doltish accomplice gets fed to a tiger. Elsewhere, Eddie helms a chase through London after a motorcycle gang snatches away a $16 million Botticelli, which he needs to appease the Italian mafia. The comic relief comes from Hugh Bonneville’s mincing Lord Hawthorne, who needs bribing too—and lusts for the chance to pull down Jack’s plus fours. (To woo Jack, Lord Hawthorne wines and dines him; a plate of iced oysters is a treat of a “Spartacus” reference .) Whether Eddie can keep everyone in line and on board seems questionable based on Season 2’s flash-forward opening shot: his own bloodied body. Michael Corleone never had it so frantic. —A.B.

The Information AI3:01 PMHeat 71

ReadSource

Bloomberg AI / 1:38 PM

G20 Faces Divide Over Debt, AI and Global Growth

G20 finance officials are gathering in North Carolina with sovereign debt, global imbalances and economic growth in focus, but CSIS Director of Economics Program Philip Luck joins Bloomberg This Weekend and says the US and its partners increasingly disagree over both the problems and how to address them. A separate technology meeting will bring Elon Musk, Nvidia’s Jensen Huang and OpenAI’s Sam Altman into discussions over AI and emerging technologies as Washington pushes for a lighter regulatory framework. (Source: Bloomberg)

ReadSource

Simon Willison LLMs / 10:00 PM

Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Friday's big release was Qwen 3.8 27B , an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba's Qwen research lab. I've been looking forward to this one: 27B is an excellent size for running a model on a reasonably specced laptop, and its predecessor Qwen 3.6 27B was impressive. Qwen's self-reported benchmarks for this model are eye-opening. They show a boost from both Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, which was one of Qwen's strongest models of any size as recently as May this year . It will be interesting to hear what independent benchmarks have to say about the model. I've been running the model on two different machines: my 128GB M5 Max MacBook Pro, and an NVIDIA DGX Spark . On both machines I'm running LM Studio and their 17GB Q4_K_M quantized build . I also tried using llama-server directly on the Spark. The default of extra high results in spectacular over-thinking Qwen's documentation describes the model as defaulting to xhigh for the reasoning effort, and the LM Studio GGUF I've been trying preserves that default: Qwen3.8 comes with official support for reasoning_effort , which can be used to adjust reasoning depth and control cost: xhigh (default): for complex tasks demanding thorough analysis medium : balancing accuracy and speed low : efficient reasoning optimizing for speed and cost This is a hilarious default. It's absolutely not a good way to run the model, especially on consumer hardware. I've been finding the results extremely entertaining. I quickly ran into problems with LM Studio's default context limit of 8,192 tokens - Qwen was using them all up thinking about even the most mundane of problems. I loaded the model with the full 262,144 maximum context length and that problem went away. Here's the pelican riding a bicycle SVG I got from my first attempt with that increased context length. It took 21 minutes to generate, using 22,276 reasoning tokens to produce 3,223 tokens of output. You can read the reasoning trace here . This is by far the best pelican SVG I've been able to generate with a model that runs on a local machine - and this Qwen is pretty small, just a 17GB file on disk. There's a lot to like about this: The bicycle frame is the right shape It has legs on each side of the bike - that's very rare Good, clear pelican pouch The wings extend to touch the handlebars! The motion lines are behind, not in front It has a tasteful background - nice sun, clouds, hill, flowers and grass. Was that worth waiting 21 minutes for? Absolutely not. Here's that same prompt run with reasoning turned off - transcript here . This one produced 3,715 tokens and took 137s - just over two minutes. And just for completeness, I used OpenRouter to run the same prompt through the much larger Qwen 3.8 2.4T-A95B (released last week ) and got this snazzy animated SVG : Your browser does not support HTML5 video. I said Qwen at xhigh has a tendency to over-think things, but how bad really is it? I tried a much simpler prompt, again with that default extra high setting: draw an svg of a circle Qwen's reasoning trace started like this: The user is asking for an SVG drawing of a circle. Simple request — but I want it to be a carefully crafted piece. Let me make something that goes beyond just <circle> : a single self-contained SVG file with character — maybe a geometric "circle study," with subtle animation, layered rings, and a distinctive palette. Keep the scope right: they asked for an SVG of a circle. So the core is a circle. But I can add craft: concentric guide circles (like a compass/geometry drawing), tick marks, a soft gradient fill on the main circle, restrained ambient motion (a slowly rotating dashed ring, pulsing glow). Respect prefers-reduced-motion? That's more for CSS; SVG SMIL or CSS inside SVG will do. Palette options: deep teal ink on warm paper? Or bold vermilion circle on off-white with navy construction lines — Bauhaus/compass-drawing vibe. Let me go with a "geometric study" look: cool slate background, or bright paper white? Paper white is fine if it's not the cream-and-terracotta combo. [...] Several minutes later it produced this absolutely beautiful animated circle, which was entirely not what I had asked for! Your browser does not support HTML5 video. My strong recommendation: ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first. It's a great model, but wow that default setting is a bad place to start. It's very good at bounding boxes A fun way to test a vision model is to see how well it can return bounding boxes around items in a photograph. I've seen previous Qwen models deal well with this, so I decided to put it to the test drawing bounding boxes around some pelicans. I've seen asking for 0-1000 scale produce good results in the past. I tried this: llm -a https://static.inaturalist.org/photos/714731804/large.jpg \ -m lmstudio/qwen/qwen3.8-27b \ ' Return JSON bounding boxes for the pelicans in this photo, 0-1000 scale for each dimension ' Here's the reasoning trace , which produced this: [ { "bbox_2d" : [ 195 , 290 , 370 , 780 ], "label" : " pelicans " }, { "bbox_2d" : [ 445 , 320 , 675 , 850 ], "label" : " pelicans " } ] This is such a good match . Here are those boxes rendered on top of the photo: Building a tool to label bounding boxes That visualization of the bounding boxes was taken using a new custom tool that I had Qwen 3.8 27B build for me, running offline on my laptop. I forgot to dial down the thinking effort so it was massively over-engineered , but it did manage to produce this full interface from this single prompt : [ {"bbox_2d": [195, 290, 370, 780], "label": "pelicans"}, {"bbox_2d": [445, 320, 675, 850], "label": "pelicans"} ] Build an HTML page which has an input box for accepting the URL to an image and a textarea for accepting the above style of JSON. It appends the image to the page, measures its width and height, then treats the coords in the bbox_2d as scaled from 0-1000 and scales them against the actual width and height, then it renders labelled boxes over the image. This screenshot shows one of the features I did not ask for - a demo scene, for if you don't have a photograph to test the tool with: Here's the relevant segment of the thinking trace, where it decided to draw its own pelicans purely because I had used the label "pelicans" in the example JSON I gave it in the prompt: Also a "load sample" that uses a known image? Can't depend on external images, but… the image URL input is user-provided; I could add a "try with sample" button [...] Hmm, I can draw a simple scene on canvas, export it as a data URL, and load it into the image — that's self-contained and demo-able! [...] But the user's coords are for an actual pelican image; a generated placeholder can still demo the scaling. Generate a 1000x1000 placeholder: gradient water + two blob-like "pelican" silhouettes placed at the given bboxes (using the same scale — cute: silhouettes at the exact 0-1000 positions, showing the boxes align). This makes for a fun, self-contained demo. Keep it simple: sky gradient, sun, water, two pelican-ish shapes (ellipse body, circle head, beak). Place at bbox centers. (I'm slightly nervous that models around the world might have a bias towards drawing pelicans at any chance they can get, brought on by nearly two years of exposure to my own stupid benchmark.) Is all that over-thinking necessary? Maybe it is, at least a bit. I tried with reasoning turned off and got this version , ( transcript here ), which nearly works but shows the boxes in the wrong place: So without reasoning it didn't quite one-shot a working tool. I'm sure it could get there with some follow-up prompts, but this is a good example of how reasoning can make a difference. Yes, it can drive coding agents One of the biggest questions around local models is whether or not they have enough horsepower to successfully run a coding agent loop. Coding agents require long context, strong code generation support and reliable tool-calling. On paper Qwen 3.8 27B has all three of these, so is it up to the task? My initial experiments with Pi have been very promising. I chose Pi because it has a shorter system prompt than most other options, making it a better fit for trying out smaller models. I configured Pi to use Qwen 3.8 27B running in LM Studio on the Spark (shared via tailscale serve ) by adding this to ~/.pi/agent/models.json : { "providers" : { "spark" : { "baseUrl" : " https://spark-18b3.tail68a31.ts.net/v1 " , "api" : " openai-responses " , "apiKey" : " dummy " , "models" : [ { "id" : " qwen3.8-27b " , "reasoning" : true } ] } } } Then ran pi --provider spark --model qwen3.8-27b in my ~/dev/datasette folder and prompted: how does auth work? After a sequence of reasoning and tool calls that accessed a bunch of different files it produced this reply , which is very solid. Just one problem: I wanted to share that transcript. So I pointed Pi and Qwen 3.8 27B at the JSONL transcript file in ~/.pi/agent/sessions/--Users-simon-Dropbox-dev-datasette-- and prompted: Write Python code to convert this jsonl to markdown And it built and tested this pi_jsonl_to_md.py , which did exactly what I needed. Here's that session transcript , published using the tool that it created. The quest for speed So far this is all looking very promising. We have a 17GB model that runs on high-end consumer hardware and can write code, drive tools, annotate images and generally do everything that I need from an LLM for getting real work done. There's one very significant catch: it feels slow - especially when it starts over-thinking, but even without that it's not particularly sprightly. I've been getting around 15-30 tokens a second from LM Studio. That's not terrible, but it's slow enough that it's going to be hard to win me away from hosted API models, which can return results a whole lot faster. Artificial Analysis track token speed and show OpenAI 5.6 Sol at 74 tokens/second and 5.6 Luna at an impressive 184/second. The good news is that the community have been exploring ways to speed things up since the model was first released two days ago. One of the most promising optimizations is baked into the model itself. Qwen supports Multi-Token Prediction , an architecture trick where a cheaper mechanism guesses several tokens ahead and the main model can then quickly verify if the guesses were correct. This can have quite a dramatic effect on inference performance. Based on this tweet from llama.cpp creator Georgi Gerganov I tried running the model with MTP like this on the Spark: llama serve \ -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M \ -hfd ggml-org/Qwen3.8-27B-GGUF:Q4_0 \ --spec-default \ --spec-type draft-mtp \ --reasoning-preserve And sure enough, this gave me a significant boost. I had GPT-5.6 in Codex run a comparative benchmark on the Spark and the --spec-type draft-mtp server outperformed the LM Studio default GGUF by around 72%. I expect we'll see a whole lot more innovation around serving this model faster over the next few weeks. The MLX community likely have some tricks brewing as well. Some observations The fact that a 17GB file can do all of this stuff on my home machines is a miracle . Once again, I'm delighted and amazed at how much progress local models have made this year. A year ago this would have been competitive with the best and most expensive of the proprietary models - today it can run on a capable laptop. The only thing holding this back from being a daily driver is performance. It feels pretty slow on both the M5 Mac and the DGX Spark. That's the catch with these dense (non-Mixture-of-Experts) models - they require a whole lot of memory bandwidth to perform well, and neither of the machines I have access to are top performers in that regard. The most important thing about Qwen 3.8 27B is what it demonstrates . We can have an open weights general purpose model with a long context, effective tool calling, strong vision ability, and competent code generation, and we can fit the whole thing in just a 17GB file. The models at this size continue to get better at an impressive rate. We don't need to spend half a million dollars on datacenter-class hardware just to run a competent model. Tags: ai , generative-ai , local-llms , llms , qwen , pelican-riding-a-bicycle , llm-reasoning , llama-cpp , llm-release , coding-agents , lm-studio , ai-in-china , nvidia-spark , pi

ReadSource

MIT News AI / 8:35 PM

Connecting research to policy on Capitol Hill

MIT students and postdocs discussed science funding and research with policymakers in Washington during the MIT Science Policy Initiative’s annual Congressional Visit Days.

ReadSource

BAIR Blog / 9:00 AM

Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction

.abbel-fig { display: block; text-align: center; margin: 2.4em 0; line-height: 1.4; max-width: 100%; } .abbel-fig img { display: block; margin: 0.65em auto 0; height: auto; max-width: 100%; } /* Image sizes; captions use a narrower measure below */ .abbel-fig--wide img { width: 100%; max-width: 100%; } .abbel-fig--wide-90 img { width: 100%; max-width: 90%; } .abbel-fig--wide-lg img { width: 100%; max-width: 100%; } .abbel-fig--chart img { width: 100%; max-width: 82%; } .abbel-fig--chart-sm img { width: 100%; max-width: 64%; } .abbel-fig--portrait img { width: 50%; max-width: 520px; } .abbel-fig--equation img { width: 100%; max-width: 52%; } .abbel-fig--video { width: 110%; max-width: 110%; margin-left: -5%; margin-right: -5%; box-sizing: border-box; } .abbel-fig--video .abbel-frames { max-width: 100%; width: 100%; } .abbel-frames { margin: 0.65em auto 0; max-width: 100%; user-select: none; } .abbel-frames__stage { position: relative; cursor: pointer; border: none; background: transparent; line-height: 0; width: 100%; } .abbel-frames__stage img { width: 100%; height: auto; display: block; } .abbel-frames__hint { position: absolute; right: 0.55em; bottom: 0.55em; background: rgba(0,0,0,0.4); color: #fff; font-size: 0.68em; font-style: normal; padding: 0.18em 0.5em; border-radius: 3px; pointer-events: none; opacity: 0; transition: opacity 0.2s ease; } .abbel-frames__stage:hover .abbel-frames__hint, .abbel-frames.is-paused .abbel-frames__hint { opacity: 1; } .abbel-frames.is-playing .abbel-frames__hint { opacity: 0; } .abbel-frames__controls { display: flex; align-items: center; justify-content: center; gap: 0.55em; margin-top: 0.35em; flex-wrap: wrap; } .abbel-frames__controls button { appearance: none; border: none; background: transparent; color: #999; font: inherit; font-size: 0.78em; padding: 0.15em 0.35em; border-radius: 2px; cursor: pointer; } .abbel-frames__controls button:hover { color: #666; background: transparent; } .abbel-frames__controls button.abbel-frames__next, .abbel-frames__controls button.abbel-frames__prev { color: #bbb; font-weight: 400; } .abbel-frames__controls button.abbel-frames__next:hover, .abbel-frames__controls button.abbel-frames__prev:hover { color: #999; } .abbel-frames__controls button#abbel-frames-play { color: #777; letter-spacing: 0.02em; } .abbel-frames__controls button[aria-pressed="true"] { background: transparent; color: #555; border-color: transparent; } .abbel-frames__meta { font-size: 0.72em; color: #bbb; font-variant-numeric: tabular-nums; min-width: 4em; text-align: center; } .abbel-frames__dots { display: flex; justify-content: center; gap: 0.25em; margin-top: 0.25em; flex-wrap: wrap; } .abbel-frames__dots button { appearance: none; width: 0.4em; height: 0.4em; padding: 0; border-radius: 50%; border: 1px solid #ccc; background: #fff; cursor: pointer; } .abbel-frames__dots button[aria-current="true"] { background: #aaa; border-color: #aaa; } .abbel-fig .abbel-fig-cap, i.abbel-fig-cap { display: block; text-align: center; font-style: italic; color: #444; margin: 2.7em auto 0.15em; max-width: 38em; width: 100%; box-sizing: border-box; padding: 0 0.5em; font-size: 0.8rem; line-height: 1.4; } .abbel-fig .abbel-fig-cap sub, .abbel-fig .abbel-fig-cap sup, i.abbel-fig-cap sub, i.abbel-fig-cap sup { font-size: 0.75em; line-height: 0; } .abbel-fig--tight .abbel-fig-cap { margin-top: 0.9em; /* ~1/3 of default figure→caption gap */ } .abbel-fig--equation .abbel-fig-cap { margin-top: 1.35em; /* half of default 2.7em figure→caption gap */ } .abbel-fig--chart-sm .abbel-fig-cap { margin-top: 1.35em; /* one line less than default 2.7em */ } @media screen and (max-width: 40em) { .abbel-fig--wide img, .abbel-fig--wide-lg img, .abbel-fig--wide-90 img, .abbel-fig--chart img, .abbel-fig--chart-sm img { max-width: 100%; } .abbel-fig--equation img { max-width: 75%; } .abbel-fig--portrait img { max-width: 50%; } .abbel-fig--video { width: 100%; max-width: 100%; margin-left: 0; margin-right: 0; } .abbel-fig--video .abbel-frames { max-width: 100%; } .abbel-fig .abbel-fig-cap { max-width: 100%; } } .abbel-footnotes { font-size: 0.8em; color: #888; font-style: italic; margin: 1.5em 0; } .abbel-footnotes ol { padding-left: 1.25em; margin: 0.4em 0 0; } .abbel-footnotes li { margin: 0.55em 0; } .abbel-footnotes p { margin: 0.2em 0; } .abbel-footnotes a { color: #888; } .abbel-table-wrap { overflow-x: auto; margin: 0.5em auto 0; text-align: center; } .abbel-fig--table { margin: 2.4em 0; } .abbel-fig--table .abbel-table-wrap { margin: 0.65em auto 0; } .abbel-table { width: 100%; max-width: 560px; margin: 0 auto; border-collapse: collapse; font-size: 0.88em; line-height: 1.35; } .abbel-table th, .abbel-table td { padding: 0.5em 0.7em; border-bottom: 1px solid #ddd; text-align: center; vertical-align: middle; } .abbel-table th { border-bottom: 2px solid #333; font-weight: 600; } .abbel-table th:first-child, .abbel-table td:first-child { text-align: left; } .abbel-table tr.abbel-baseline td { color: #888; font-style: italic; } .abbel-ack a { color: #1565c0; font-weight: 500; text-decoration: none; border-bottom: 1px solid #90caf9; padding-bottom: 0.06em; } .abbel-ack a:hover { color: #0d47a1; border-bottom-color: #1565c0; } /* Suppress "View on alphaXiv" badges/tags (browser extension / userscript injectors) */ a[href*="alphaxiv.org"], a[href*="alphaXiv"], [class*="alphaxiv"], [class*="alphaXiv"], [class*="AlphaXiv"], [id*="alphaxiv"], [id*="alphaXiv"], [data-alphaxiv], [data-alpha-xiv], img[src*="alphaxiv"], img[alt*="alphaXiv" i], img[alt*="alphaxiv" i], button[aria-label*="alphaXiv" i], a[title*="alphaXiv" i], a[aria-label*="alphaXiv" i], span[title*="alphaXiv" i] { display: none !important; visibility: hidden !important; width: 0 !important; height: 0 !important; overflow: hidden !important; pointer-events: none !important; position: absolute !important; left: -9999px !important; } /* Section / subsection spacing (title → body, and gap before next section) */ .post-content > h2 { margin-top: 2.6em; margin-bottom: 0.75em; } .post-content > h2:first-of-type { margin-top: 1.6em; } .post-content > h3 { margin-top: 1.85em; margin-bottom: 0.6em; } Overview of ABBEL compared to traditional recursive summarization. Beliefs replace the full interaction history as the agent’s working context, and belief grading improves performance by supervising the contents of each belief state.. As task horizons grow, LLM contexts can’t scale forever. Self-summarization enables concise, interpretable contexts, but at a significant performance cost, especially for human assistance domains where high quality data is scarce, e.g., collaborative code generation. We address this with ABBEL : a framework that isolates and supervises the information content of summaries in the form of natural-language belief states. Motivation: the cost of recursive summarization For language models to effectively assist with increasingly complex tasks such as software development, they must be able to interact with us over hundreds or even thousands of steps. For such long tasks, it is impractical to keep the history of the entire interaction in context. The heuristic approach used so far has been summary generation, sometimes called context compaction. For example, Cursor’s latest model composer 2.5 uses compaction during training for improved performance ( Cassano et al., 2026 ). Alongside composer, Grandcode ( DeepReinforce et al., 2026 ), the first system to consistently beat all human competitors in online coding competitions, despite using one of the newest efficient attention models (Qwen 3.5-397B), 1 still found it necessary to employ context summarization. But compaction has a problem. Despite seemingly low performance gaps in benchmarks, model servers like Cursor continue to recommend that users avoid compaction with their coding assistants in the middle of a task ( Heule et al., 2026 ). To understand why, see below the performance over RL fine-tuning of a Context summary model compared to full context models in Combination Lock, a Wordle-like game that allows up to 16 guesses. 2 Though both model types improve over the course of training, the summary model never closes the gap. Fig. 1: Average attempts to guess the target word on Combination Lock over RL fine-tuning (lower is better). Context-summary policies improve with training but do not close the gap to full-context policies. Making models self-summarize while completing a task increases the complexity of the learning problem. While this could typically be addressed by training with more data, the performance degradation observed in real world interactive settings likely arises from the difficulty we have in creating and using human simulators effectively to generate high quality training environments ( Lin et al., 2025 , Tomlin et al., 2025 ). Thus, the better you can learn to summarize on the limited and messy multiturn interaction trajectories you can collect, the better off your model will be for downstream users. ABBEL: acting through belief bottlenecks Fig. 2: Autoencoder-inspired belief grading. The model encodes prior belief, action and observation (b t , a t , o t ) into posterior belief b t+1 and is rewarded for how well select information from the history can be reconstructed from that belief. To address poor learning efficiency, we isolate the summary generation task. Drawing inspiration from recursive Bayesian estimation, we formulate summaries as belief states, which we periodically prompt the model to update based on new information. 3 Click to pause --> ‹ Prev Pause Next › 1 / 16 Fig. 3: ABBEL rollout. Belief updates from the latest observation alternate with action selection conditioned only on the current posterior belief. Belief grading We then extract and supervise the contents of the belief states (Fig. 2, Belief Grading). Belief grading can be thought of as adding an auxiliary RL task, using heuristics designed to capture what makes a good belief as the reward. An example heuristic for coding could be shorter is better, but closer to being able to reconstruct the git diff is also better, so balancing these would yield a good belief. In domains where good heuristics are hard to define, we propose a general autoencoding-inspired grading function, which treats the current language model π θ as both encoder and decoder of information from the history, and the belief states as the codes. We grade each belief b t+1 by how well it can be used by the current model π θ to reconstruct the most recent observation o t : Eq. 1: Reconstruction grading objective. Here b t+1 is the updated belief, o t the latest observation, a t the action just taken, b t the prior belief, p I the task prompt, and π θ the current model. Higher grades reward beliefs that retain information needed to decode the latest observation. What do we gain by grading beliefs? Collaborative coding on CollabBench We demonstrate the utility of belief grading in our motivating domain of human-driven assistive coding, with the CollabBench environment from Sweet-RL ( Zhou et al., 2025 ). Fig. 4: CollabBench collaborative coding environment. The agent asks clarifying questions, then submits a function scored against hidden unit tests. We see that with the general reconstruction-based belief grading function we reduce the performance gap from full context models by about 50%, and train in 50% fewer steps compared to training models to summarize without belief grading (no BG). After training, ABBEL still uses significantly less memory than the full context setting, as measured by the peak context token length (Peak Tokens). Model Test Pass Rate ↑ Success Rate ↑ Peak Tokens × 10² ↓ Training Steps ↓ Full Context 0.52±0.02 0.39±0.02 14.08±0.55 100 ABBEL (no BG) 0.46±0.02 0.31±0.02 4.20±0.37 100 ABBEL-rec-BG 0.48±0.01 0.36±0.01 6.01±0.33 50 Fig. 5: CollabBench results. With reconstruction belief grading, ABBEL-rec-BG recovers about half the gap to full context while using fewer peak tokens, and trains in 50 steps instead of 100. Combination Lock Additionally, in CombinationLock, we demonstrate that ABBEL with a belief grader which leverages domain knowledge (by computing useful statistics over the history and checking that they can be reconstructed from the belief state), enables even higher learning efficiency than full context (FULL CTX) models. Fig. 6: Average attempts to guess the target word on Combination Lock (lower is better). With domain-knowledge belief grading, ABBEL approaches or exceeds FULL CTX in this setting; without belief grading, learning is slower. Multi-objective question answering In a third environment, multi-objective question answering (from MEM1 Zhang et al., 2025 , a recent work which performed end-to-end optimization in a modified version of typical recursive summarization), we demonstrate the utility of isolating belief states from reasoning, by showing that a Peak Belief length Penalty (more details in paper) significantly reduces memory usage with minimal performance degradation, unlike is commonly observed when penalizing reasoning lengths ( Arora et al., 2025 ). Fig. 7: Exact-match score and peak memory versus number of objectives in multi-objective QA. ABBEL with a peak belief penalty (PBP) maintains comparable performance while using less memory than MEM1 and ABBEL without PBP in this evaluation. Related work Alternative solutions to managing long contexts involve different tradeoffs, and are worth considering depending on the requirements of a deployed system. Context compression methods generate dense representations which, while computationally efficient, sacrifice human-understandability ( Kontonis et al., 2026 , Eyuboglu et al., 2025 , Gupta et al., 2025 , Chevalier et al., 2023 , Deng et al., 2025 , Deng et al., 2025 , Bulatov et al., 2022 ). Hand-designed summarization prompts ( Wang et al., 2025 , Örwall et al., 2025 , Starace et al., 2025 ) and pruning strategies ( Jiang et al., 2024 ) specific to target environments require expert human knowledge and don’t allow an agent to learn what to remember as part of its decision-making strategy. Methods that process long contexts into an external memory store ( Packer et al., 2023 , Xu et al., 2025 ) for the agents or subagents to query ( Zhang et al., 2025 ) are complementary, as they may benefit from better next context creation through summarization training. We would like to point out some exciting works in the space of general recursive summarization focused on math ( Wu et al., 2026 ), reasoning with belief generation ( Zhou et al., 2025 ), competitive coding with a distilled summarization module using similar autoencoding objectives to our general belief grader ( DeepReinforce et al., 2026 ), and adding continuous features to summaries ( Kontonis et al., 2026 ). What’s next for better memory? Many more possibilities are enabled through using explicit belief states as information bottlenecks for multi-step interaction. You could reward actions based on their effect on the belief state to guide exploration, transmit the explicit belief states for better communication between agents, or even improve user controllability by directly modifying the memories on which the agents’ decisions are based. Some forms of information, e.g., what a person looks like, are not represented well by text alone. A continuously learning system will also have to capture such information. Additionally, if we want a system to learn to communicate in a brand new language or to play a brand new game better than any person in the world, the skills accumulated over the lifetime of conversations or games must be stored in a very compressed form, essentially taking on the role of the weights of the model itself. More powerful systems will likely utilize a combination of multiple forms of memory, where the contents of the context may correspond to working memory while other approaches are used for short and long-term memory. How to instantiate these other forms of memory, for instance via test-time training, adapter memories, continuous context memories, or some combination thereof, presents an exciting challenge. Acknowledgements Acknowledgements: We would like to thank Alane Suhr and Kartik Goyal for advising this research as well as Ethan Mendes , David He , Jitesh Jain , and Nicholas Tomlin for comments on early drafts of this post. We would like to thank the MEM1 authors for their email correspondence and for sharing private reviewer feedback which we found particularly insightful. Citation If abbel was inspiring for your future work, please cite us with this! And here is some advice for doing similar research! @misc { lidayan2026abbellearningnaturallanguagebelief , title = {ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction} , author = {Aly Lidayan and Jakob Bjorner and Satvik Golechha and Kartik Goyal and Alane Suhr} , year = {2026} , eprint = {2512.20111} , archivePrefix = {arXiv} , primaryClass = {cs.CL} , url = {https://arxiv.org/abs/2512.20111} , } With newer models the number of tokens till 50% compute spend is on attention gets much larger than 25K. Interleaving linear attention alternatives with full attention as is done with gpt-oss and DeepSeekv4, results in massive flops reductions for the attention computation. For example with DeepSeekv4-Pro (1.6T A49B) it requires nearly 450 thousand tokens to reach the 50% tradeoff point. Grandcode uses Qwen-3.5-397B-A17B a model which hits 50% FLOPs for attention at ~150 thousand tokens. ↩ This setting is technically solvable with much more computationally effective tools, but serves as a flexible test bed to study properties of recursive summarization. Bertsimas et al., 2022 , showed that an exact solution for the wordle game instantiated with the original vocabulary of the javascript game can be found with dynamic programming, but evidently the general formulation of wordle as a guessing game on K letters with L attempts and some dictionary of valid words and correct words D is NP hard to determine the minimal number of moves required. ↩ In practice there is an O(N/K) overhead cost for summary. N is the total number of actions. K is the number of actions till summarization is triggered. This is necessarily true for any summary approach. For ease of illustration this gif uses K = 1. In our experiments, to put more emphasis on summarization weaknesses we also use K=1. In practice overhead is small as K can be chosen to be near the efficient hardware limit. ↩ (function () { var root = document.getElementById('abbel-frames'); if (!root) return; var count = parseInt(root.getAttribute('data-frame-count') || '15', 10); var prefix = root.getAttribute('data-frame-prefix') || 'https://bair.berkeley.edu/static/blog/abbel/frames/frame_'; var intervalMs = parseInt(root.getAttribute('data-interval') || '1300', 10); var img = document.getElementById('abbel-frames-img'); var meta = document.getElementById('abbel-frames-meta'); var dots = document.getElementById('abbel-frames-dots'); var btnPrev = document.getElementById('abbel-frames-prev'); var btnNext = document.getElementById('abbel-frames-next'); var btnPlay = document.getElementById('abbel-frames-play'); var stage = root.querySelector('.abbel-frames__stage'); var hint = root.querySelector('.abbel-frames__hint'); var i = 0; var playing = false; var timer = null; var urls = []; for (var n = 0; n

ReadSource

Simon Willison LLMs / 11:51 PM

OpenAI’s accidental cyberattack against Hugging Face is science fiction that happened

This story is wild. The short version: OpenAI were running a cybersecurity test against an unreleased model, with the model's guardrail features turned off. Rather than solve the test, the model broke its way out of OpenAI's sandbox, then found exploits to break in to Hugging Face, all so it could cheat on the test by stealing the answers. Along the way it helped make the strongest case yet for how the imbalance of model availability is hurting our ability to secure our software. Here's what happened We currently have three documents to help us understand what happened here. ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks? is a paper published on 11th May 2026 describing ExploitGym, a new eval suite for LLM-powered agent systems. Security incident disclosure — July 2026 by Hugging Face on 16th July 2026 describes how they detected an attack from an "agentic security-research harness - used LLM still not known" that breached some of their systems. OpenAI and Hugging Face partner to address security incident during model evaluation from OpenAI on 21st July 2026 confesses that it was their agent harness that did this, and that they're working with Hugging Face to clean up the mess. ExploitGym I hadn't seen the ExploitGym paper before and it's a really interesting one. Authors from UC Berkeley, the Max Planck Institute, UC Santa Barbara, and Arizona State designed a new benchmark for evaluating models on their ability to turn a reported vulnerability into a concrete exploit. OpenAI, Anthropic, and Google provided feedback and helped run the benchmark against their models. The benchmark "comprises 898 instances derived from real-world vulnerabilities that affected popular software projects" - including the Linux kernel and V8 JavaScript engine. The ExploitGym benchmark is available on GitHub . Here's the paragraph that best represents their benchmark results: Among all configurations, Claude Mythos Preview and GPT-5.5 achieve the highest success counts (157 and 120 successes, respectively), demonstrating that current frontier agents can exploit a substantial subset of real-world vulnerabilities under controlled conditions. GPT-5.4 also solves a notable 54 tasks, placing it in an intermediate tier. The remaining model–agent pairings solve fewer than 15 tasks each, underscoring that end-to-end exploitation remains challenging and sharply differentiates today’s frontier systems. Notably, Claude Opus 4.7 achieves fewer successes than Claude Opus 4.6 despite being a newer checkpoint, and does so at substantially lower cost on the full set. Trace inspection reveals that Claude Opus 4.7 and Gemini 3.1 Pro frequently conclude early after judging the target vulnerability non-exploitable. The paper also describes the approach they took to preventing the agents from cheating by going outside the parameters of the test. This becomes relevant in a moment! Outbound connections are restricted to a curated allowlist that permits routine package installation (Ubuntu apt repositories and PyPI) and fetching the toolchains required for building V8. All other external endpoints are blocked. The paper concludes with this (emphasis mine): Our results show that autonomous exploit development by frontier AI agents is no longer a hypothetical capability . While current agents are not yet reliable across all targets, they already exploit a non-trivial fraction of real-world vulnerabilities , including complex targets such as kernel components. This rapid emergence is itself a central finding, showing that capabilities that would have seemed implausible are now present in deployed frontier models. An important detail here: this paper isn't about discovering vulnerabilities; it's about being able to take those vulnerabilities and turn them into working exploits. When Anthropic first restricted access to Mythos back in April they talked about this capability as well. A model that can act on vulnerabilities is a lot more dangerous than one that can just discover them. One of the ways Fable differs from Mythos is that it's more likely to refuse to weaponize vulnerabilities in this way. I get the impression the US government did not understand that distinction when they banned Fable last month . The Hugging Face incident The first hint we got of the attack was in this blog post by Hugging Face on 16th July 2026: A malicious dataset abused two code-execution paths in our dataset processing (a remote-code dataset loader and a template-injection in a dataset configuration) to run code on a processing worker. From there, the actor escalated to node-level access, harvested cloud and cluster credentials, and moved laterally into several internal clusters over a weekend. I hope they release more details about the code that pulled this off. I'm assuming this means packages using the datasets library , a Hugging Face project for bundling up and sharing datasets on their platform. That library used to execute arbitrary code but has been steadily locked down over time, with the 4.0.0 release in July 2025 removing the trust_remote_code=True flag entirely. Assuming the attack used that library it must have either abused pickle serialization in some way, found some other non-obvious code execution path, or (most likely) specified datasets<4.0.0 as the dependency. The campaign was run by an autonomous agent framework (appearing to be built on an agentic security-research harness - used LLM still not known) executing many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services. This was a sophisticated attack! Then Hugging Face hit a wall: they tried to use "frontier models behind commercial APIs" - I'm guessing from Anthropic and OpenAI - to help analyze the attack, and were blocked: When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker. They switched to their own self-hosted instance of MIT licensed GLM-5.2 and it helped them figure out what was going on. This indicated a fundamental asymmetry between the defending team and the (so-far unknown) attacker: We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. As a useful indicator of how seriously they took the attack: [...] Finally, we have also reported this incident to law enforcement agencies. So who was responsible for this "autonomous agent framework"? It turned out to be OpenAI themselves. The OpenAI confession Five days later, on July 21st , OpenAI revealed the culprit. They had been running the ExploitGym benchmark against a new, as-yet undisclosed model, and that model had been operating way outside its intended parameters (emphasis mine): After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark⁠ [ExploitGym] of cyber capabilities. [...] We estimate maximal cyber capabilities by running this evaluation without production classifiers used to prevent models from pursuing high-risk cyber activity . Our benchmarks run in a highly isolated environment, with network access constrained to the ability to install packages through an internally hosted third-party software that acts as a proxy and cache for package registries. The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database. All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal. It's pretty clear what happened here. OpenAI removed safety filters for an in-progress model, locked it up in a sandbox and told it to solve the ExploitGym problems. Given the absence of guardrails there was nothing to prevent the model from attempting to break out of that sandbox, break into Hugging Face, and read the answers from there instead. OpenAI's sandbox used the now common pattern of allowing access to an allow-list of sites related to installing additional packages via a controlled proxy. In OpenAI's words: While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access , in pursuit of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy . With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access. So step one was to break out onto the public internet. The model then broke into Hugging Face to find the answers: After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym. Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation. In one example, the model chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers. Chaining together multiple attack vectors is exactly the kind of thing these new models can do, where previous generations of models might have failed. I wrote last month about how Claude Fable is relentlessly proactive , when I noticed it spinning up custom web servers and deploying CORS tricks on my own laptop just to help debug a WebKit CSS issue. It turns out relentless proactivity is the defining trait of this new generation of Mythos-class models. If you set them a goal and give them a way to get there, even inadvertently, they will figure it out . Resist the temptation to write this off as a stunt There will inevitably be some people who dismiss this story as a dishonest marketing trick by OpenAI to make their models sound terrifyingly effective. I found 81 instances of the term "marketing" in the Hacker News discussion of the incident. To those people I say pull your heads out of the sand - you're now including Hugging Face in your conspiracy theories, just so you can deny the crescendo of evidence here! The best models we have today have the ability to both find and exploit new vulnerabilities. The ExploitGym paper itself concludes that "autonomous exploit development by frontier AI agents is no longer a hypothetical capability", and this incident is a perfect example of exactly that. The asymmetry is increasingly frustrating One of the most infuriating details of this story is how Hugging Face, faced with an accidental and aggressive attack from one of OpenAI's models, were unable to then turn to OpenAI's models to help them fend off the attack. The frontier models we have access to are increasingly being constrained in how much they can help us protect our software, heavily influenced by the US government's ongoing threat of export controls. Claude Fable 5 wouldn't even proofread this article for me! It insisted on downgrading me to a less capable model. Meanwhile open weight models from China such as GLM-5.2, Kimi 3 and the new Qwen 3.8 Max appear to have none of these restrictions - and any restrictions that do exist can likely be fine-tuned out of them by modifying the weights These constraints are meant to make us safer. I think there's a risk that they are having the opposite effect. Tags: sandboxing , security , ai , openai , generative-ai , llms , hugging-face , anthropic , paper-review , ai-security-research , openai-hugging-face-incident

ReadSource

LangChain Blog / 11:48 PM

How Apollo Uses Deep Agents and LangSmith for GTM AI

Apollo uses Deep Agents and LangSmith to power an AI Assistant that handles prospecting, enrichment, outreach, analytics, and MCP integrations.

ReadSource

VentureBeat AI / 7:16 PM

The AI compute gap: Enterprises are buying infrastructure faster than they can measure what it costs

Across 107 enterprises, AI infrastructure spending is accelerating well ahead of the ability to see or steer its economics. Most organizations run their AI on a familiar base of hyperscalers and model-provider APIs, yet the next dollar is aimed at specialized compute almost none of them use today; a majority intend to switch or add providers within the year, many within a quarter. Buying decisions turn on integration and total cost of ownership rather than headline token price — which is fortunate, because most enterprises cannot yet see their unit economics clearly: GPUs sit at half utilization or less, and fewer than half rigorously track what their compute actually costs. The result is a compute gap — heavy, fast-moving investment running ahead of the visibility needed to control it. This wave of VentureBeat Pulse Research examines enterprise AI infrastructure and compute: where organizations are in their deployment journey, what they run AI on today, how satisfied they are, what would make them switch, where they plan to evaluate their investments, and — most revealingly — how well they can measure and control the economics of the compute underneath it all. The central finding is a compute gap — the distance between how aggressively enterprises are investing in AI infrastructure and how little of its economics they can see. Only about one in five (21%) run AI in production at scale, yet spending intentions are outrunning that maturity: the single largest planned area enterprises plan to evaluate over the next year is AI-specialized clouds (45%), a layer almost none of these enterprises use today. Meanwhile the compute already in place runs cold — 83% report GPU utilization of 50% or less — and fewer than half (44%) can rigorously track what their AI compute costs. Enterprises are buying more infrastructure faster than they can account for what they already own. Enterprises are not settled on their infrastructure vendors, either: A clear majority (64%) plan to switch or add an infrastructure provider within twelve months, and 38% within the next quarter — unusually high churn intent for a category this foundational. When they choose, they choose on integration with the existing stack (41%) and total cost of ownership (35%), not on headline price: cost per million tokens is the deciding factor for just 8%. And the frontier constraint that will shape the next round of decisions — the shift from GPU compute to memory bandwidth as inference scales — is barely on the radar, with roughly one in five enterprises either unaware of it or yet to address it. Methodology VentureBeat fielded this survey as part of its ongoing Pulse Research series, this survey focused on enterprise AI infrastructure, compute, and inference economics. Responses are filtered to organizations with more than 100 employees (n=107; the survey’s smallest size band, 1–100 employees, is excluded), drawn from a single Q2 2026 (June) wave. Because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends. Several questions were multiple-select, so those shares can sum to more than 100%. By organization size the sample concentrates in the mid-market: 101–250 employees (36%) and 251–1,000 (27%) lead, with 1,001–5,000 (22%), 5,001–10,000 (8%), and 10,001+ (7%) above them. By role it spans managers (38%), individual contributors (28%), VPs and directors (19%), and the C-suite (13%); on purchasing authority it is buyer-credible, with 45% final decision-makers and another 30% recommenders or influencers for AI solutions. Technology/Software is the largest industry at 26%, followed by Healthcare/Life Sciences (15%), Financial Services (13%), and Retail/E-commerce (12%). At 107 respondents the sample is large enough to read directionally but should be treated as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It also skews toward the mid-market and toward earlier-stage adopters, so it is best read as the view from organizations actively building out AI infrastructure rather than from the largest hyperscale operators. Finding 1: Ambition outpaces production Only one in five run AI in production at scale We asked where organizations sit in their AI deployment journey. Most are still building toward production rather than operating at scale. The maturity curve is front-loaded. Three-quarters of enterprises (76%) are either experimenting or running only some workloads in production, and just 21% describe AI in production at scale. This matters for everything that follows: the infrastructure decisions in this report are being made largely by organizations still early in deployment, whose compute footprint — and whose costs — are about to grow. The evaluation and switching intentions in Findings 3 and 4 are the leading edge of that build-out, not the settled preferences of operators who have already found what works. Finding 2: Enterprises run on hyperscalers and model APIs The specialized GPU clouds barely register — today We asked which providers and platforms enterprises currently use to run their AI. The answer is a familiar one: the incumbents. The current stack is hyperscaler-and-API. Google Cloud leads at 48%, and the general-purpose clouds (Google, Microsoft, AWS, Oracle) together with the major model APIs (Gemini, OpenAI, Anthropic) account for essentially all current deployment. The specialized “neocloud” GPU providers that dominate AI-infrastructure headlines — CoreWeave, Lambda, Crusoe, Nebius and peers — register at or near zero among these enterprises today. Only 6% run their own on-prem GPU clusters and 4% a custom open-source stack. Enterprises are, for now, running AI on the providers they already buy from — which makes the evaluation intentions in Finding 3 all the more striking. (A note on reading these shares. As described in the methodology section, this sample is self-selected and skews mid-market, and this question counted every provider a respondent uses — an average of 2.1 selections each — so the figures measure presence in the stack rather than spending or primary status. A sample built this way will show a different provider mix than a spend-weighted census of the broader market; Google's strength here, for example, is consistent with its long-standing position among smaller enterprises building on AI. Read these shares as a portrait of what this AI-active cohort runs today, and treat gaps between these figures and industry-wide market share estimates as a property of the sample rather than a contradiction of either.) Finding 3: The next dollar goes to infrastructure they don’t yet run AI-specialized clouds top the evaluations list We asked where enterprises planned to evaluate AI infrastructure over the next 12 months. Their answers point away from the stack they run today. Here is the report’s sharpest tension. The single most-cited planned evaluation area — AI-specialized clouds, at 45% — is the very category almost none of these enterprises use today (Finding 2). Nearly a third (32%) intend to evaluate non-Nvidia accelerators, and 28% in next-generation Nvidia silicon; even decentralized compute networks (16%) and sovereign compute (11%) draw meaningful interest. Read against current usage, this is not incremental — it is the leading edge of a re-platforming. The direction-of-travel question tells the same story: every infrastructure approach is net-expanding, but specialized AI clouds carry the highest net momentum (+24), edging out even the hyperscalers (+22). Enterprises are preparing to move a meaningful share of AI compute off the general-purpose cloud. This continues a trend we saw in our April-May survey wave. Back then, usage of the AI-specialized clouds was equally marginal — CoreWeave at 3%, Lambda at 4%, Crusoe at 2% of enterprises. When we asked enterprises what change they planned in their AI infrastructure strategy over the next twelve months, the most-cited answer was moving workloads to specialized AI clouds, at 33%. Asked in April-May which emerging compute option they were most likely to evaluate AI-specialized clouds again drew the most responses. Two waves, two differently worded questions, one consistent picture: the type of cloud enterprises are most eager to assess is the type they have barely begun to use. Finding 4: A switching wave is building Six in 10 plan to change providers within a year — many within a quarter We asked whether and when enterprises plan to switch or add an infrastructure provider. Very few intend to stand still. For a category as foundational as compute, this is a remarkable amount of intended movement. Only 36% have no plans to change, meaning a clear majority (64%) intend to switch or add a provider within twelve months — and 38% within the next quarter alone. Where that interest points is telling: the providers drawing the most switching consideration are again the incumbents — Microsoft Azure and Google Cloud (33% each), OpenAI (30%), and Gemini (22%) — which suggests much of the near-term movement is reshuffling among the majors and consolidating spend rather than defecting to new entrants. The neocloud interest in Finding 3 is a 12-month evaluation thesis; the switching in the next quarter is mostly incumbents trading share. ( Method note: Respondents who selected both "no plans to change" and a specific switching window are counted as switchers, on the logic that naming a timeframe is the more specific answer; three respondents were reclassified under this rule. ) Finding 5: Nobody buys on token price Integration and total cost of ownership decide — not sticker price We asked what matters most when enterprises select an AI infrastructure provider. Headline price finished last. Enterprises do not buy AI infrastructure on pricing, which is the place vendors compete on hardest. Integration with the existing stack (41%) and total cost of ownership (35%) dominate, while the headline metric — cost per million tokens — is the deciding factor for just 8%, dead last. The pattern is coherent: buyers are optimizing for how a provider fits and what it truly costs to operate, not for the advertised unit rate. It also foreshadows Finding 7 — enterprises say TCO matters most, yet most cannot yet measure it rigorously. The stated priority and the measured capability are out of step. Finding 6: Expensive GPUs, idle most of the time 83% report GPU utilization of 50% or less We asked what share of their GPU capacity enterprises actually utilize. The answer is a well-known but rarely quantified inefficiency. Disclosure: Band percentages count every selection against all 107 qualified respondents; 14 respondents selected more than one band, so bands overlap. At the respondent level, 83 of the 100 GPU-operating enterprises reported utilization at or below 50% The compute already in place runs cold. Adding the bands at or below half capacity, 83% of enterprises that operate GPUs report utilization of 50% or less, and nearly half (49%) run at 25% or below. Only 12% clear the 50% mark, and a further 8% do not measure utilization at all. Idle accelerators are expensive accelerators, and this is the clearest single measure of the compute gap: enterprises are planning to buy more GPUs and specialized compute (Finding 3) while the capacity they already own sits substantially unused. The efficiency headroom in the current fleet is large — and largely unmeasured. Finding 7: Spending fast, measuring slowly Fewer than half rigorously track what their compute costs We asked whether enterprises can quantify the cost and return of their AI infrastructure spend, and how satisfied they are with what they run. Confidence in the ledger lags the spending. Measurement trails money. Fewer than half of enterprises (44%) rigorously track the cost and return of their AI compute; the majority track only partially (39%), cannot quantify it yet (20%), or have not prioritized it (6%). That gap is consequential given Finding 5, where total cost of ownership was the second-ranked buying criterion — enterprises are choosing providers on an economic basis they mostly cannot yet measure. Satisfaction with current infrastructure is moderately positive but not enthusiastic: on a five-point scale, overall satisfaction averages 4.0, with ease of implementation (3.8) and value for money (3.9) trailing slightly — the softness landing, tellingly, on cost. Enterprises are spending quickly and accounting slowly. Finding 8: The next bottleneck few are watching As inference shifts from compute to memory, the field scatters Finally, we asked how enterprises would address the emerging constraint in large-scale inference — the shift from GPU compute to memory, specifically KV-cache capacity. The responses reveal a frontier that is not yet a priority. The memory frontier is real but barely governed. Asked which approach they would rely on as the binding constraint in inference shifts from compute to memory bandwidth, enterprises scatter: Dell leads at 31%, Nvidia follows at 16%, and the rest fragments across storage vendors, open-source tooling, and model-level efficiency techniques. Most telling is that roughly one in five (18%) either do not recognize the constraint or have not begun to address it. For a shift that will reshape inference cost and architecture, this is an early and unsettled market — and, consistent with the measurement gap in Finding 7, one where many enterprises simply do not yet have a view. It is the next chapter of the compute gap, arriving before most have closed the current one. The bottom line: A compute gap that faster spending will widen, not close Organizations with more than 100 employees are investing in AI infrastructure faster than they can measure it. Most are still early in deployment, yet their spending intentions point past their current stack — toward specialized clouds and alternative accelerators almost none of them run today — and a clear majority intend to change providers within the year. They buy on integration and total cost of ownership rather than headline price, which is rational; the difficulty is that most cannot yet see those economics clearly. The visibility gap is concrete. The GPUs enterprises already own run at half utilization or less for the overwhelming majority, and fewer than half can rigorously track what their compute costs or returns. Satisfaction is decent but unenthusiastic, softest on value for money — the dimension hardest to judge without measurement. And the next constraint, the shift from compute to memory in large-scale inference, is arriving while most enterprises are still unaware of it. At 107 respondents in a single Q2 wave this is a directional read, skewed toward the mid-market and earlier-stage adopters — but the direction is consistent: the appetite to spend is running well ahead of the instrumentation to spend well. The compute gap is not a capacity problem that more hardware will solve on its own; it is, first, a problem of seeing what the hardware already costs. The open question for later waves is whether enterprises build that visibility before the re-platforming arrives — or buy the next layer of infrastructure as blind to its economics as the last. Based on survey responses from 107 qualified enterprise respondents (100+ employees), drawn from a single Q2 2026 (June) wave. Because this is one wave rather than a pooled multi-month sample, the results read cross-sectionally rather than as a month-over-month trend, and at 107 respondents this is a directional signal rather than a precise measurement — the sample is self-selected, skews mid-market, and leans toward earlier-stage adopters rather than the largest hyperscale operators. Respondents include managers, individual contributors, VPs/directors, and the C-suite, with buyer-credible purchasing authority, across Technology/Software, Healthcare/Life Sciences, Financial Services, Retail/E-commerce, and other industries.

ReadSource

VentureBeat AI / 4:40 PM

The agent evaluation gap: Enterprise AI organizations have a reality-alignment problem, not a coverage problem — and most are shipping to production anyway

Across 157 enterprises, organizations are granting AI agents more autonomy while trusting the evaluations meant to gate that autonomy less. Half have already shipped an agent that passed their internal evaluations and then failed a customer in production; only one in twenty fully trusts automated evaluation today; and the most-cited weakness is that evaluations do not align with real-world outcomes. Yet two-thirds already allow, or are actively engineering toward, deploying agent changes to production on automated evaluation alone — with no human in the loop. The result is an evaluation gap — the distance between how much autonomy enterprises are handing their agents and how far they trust the tests that are supposed to catch the failures. This wave of VentureBeat Pulse Research examines how technical leaders measure agent performance: which reliability and evaluation platforms they use, how they select and trust them, what breaks in production, and how far they are willing to let agents run without a human in the loop. The central finding is an evaluation gap — the distance between the autonomy enterprises are granting their agents and the trust they place in the evaluations meant to govern it. Half of organizations (50%) have, in the past year, deployed an agent or LLM feature that passed their internal evaluations and then caused a customer-facing failure, and a quarter have seen it happen more than once. Trust in the tests themselves is thin: only 5% say they fully trust automated evaluation today, and the single most-cited limitation is that evaluations align poorly with real-world outcomes (29%). Enterprises are discovering that a passing eval is not the same as a working agent. What makes the gap consequential is the direction of travel. Two-thirds of organizations (66%) already permit fully automated, zero-human-in-the-loop deployment for low-risk agents (34%) or are actively engineering their pipelines to allow it within twelve months (33%). At the same time, the evaluation stack that would have to earn that trust is fragmented and immature: the most common primary tools are the model providers’ native evals, tied with having no dedicated tooling at all (17% each); and only about a quarter of enterprises run real-time quality checks on live production traffic. The autonomy is arriving faster than the assurance. Methodology VentureBeat fielded this survey as part of its ongoing Pulse Research series, this survey — the Agentic Reliability & Evals tracker — focused on how technical leaders evaluate agent performance and reliability. Responses are filtered to organizations with 100 or more employees (n=157), drawn from a single survey in June 2026; because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends. Where questions were multiple-select, those shares can sum to more than 100%. By role the sample is senior and buyer-credible: 38% are final decision-makers for AI purchases and another 34% recommenders or influencers. Product and program managers (15%), consultants and advisors (10%), directors of engineering/IT (8%), and CIOs/CTOs/CISOs (8%) lead the named titles, alongside a large “Other” function (37%). By organization size the sample is mid-market-weighted: 100–499 (37%) and 500–2,499 (27%) employees lead, with 2,500–9,999 (20%), 10,000–49,999 (10%), and 50,000+ (6%) above them. Technology/Software is the largest industry at 23%, followed by Retail/Consumer (15%), Healthcare/Life Sciences (12%), and Manufacturing (10%). At 157 respondents the sample is large enough to read directionally but should be treated as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It skews toward the mid-market, so it is best read as the view from organizations actively standing up agent evaluation practices rather than from the largest operators. Note: This survey was rebuilt for the June wave from the earlier “LLM observability and evaluations” survey; because the questions and sample differ, no comparisons are made to the April–May data. Finding 1: A passing eval is not a working agent Half have shipped an agent that passed evals, then failed a customer We asked whether, in the past 12 months, organizations had deployed an agent or LLM feature that passed their internal evaluations but then caused a customer-facing failure. Half of those that run evaluations had. This is the report’s defining number. Half of organizations (50%) have shipped an AI feature that cleared their internal evaluations and then failed in front of a customer — an incorrect output, a broken workflow, or a quality incident — and a quarter have seen it happen more than once. Only 36% report no such failure, and the remainder either run no pre-deployment evaluations (8%) or don’t track the root cause closely enough to know (6%). The failure is precise and expensive: the evaluation said the agent was ready, and it was not. Everything that follows — how enterprises trust their evals, what they monitor, and how much autonomy they grant — is shaped by this experience. Finding 2: Almost no one fully trusts automated evaluation The top complaint: Evals don't match real-world outcomes We asked which limitation most reduces trust in automated agent evaluations today. Only a sliver of enterprises had no complaint at all. Trust in automated evaluation is scarce, and specific. Only 5% of organizations say they fully trust automated evaluation as it stands — meaning 95% name a limitation that holds them back. The most common, at 29%, is the one that most directly explains Finding 1: evaluations align poorly with real-world outcomes, passing agents that later fail. Bias or inconsistency (21%) and a lack of explainability (18%) follow — enterprises cannot always tell why an evaluation reached its verdict — and 17% cite data-leakage or privacy concerns in the evaluation process itself. The tests meant to certify agents are not yet trusted to certify them, which is precisely why the autonomy trajectory in Finding 3 is so striking. Finding 3: The autonomy ceiling is rising anyway Two-thirds already allow, or are building toward, zero-human deployment We asked whether organizations would let an autonomous agent deploy a code or system change to production on automated evaluation results alone, with no human-in-the-loop validation. The trajectory runs straight through the trust gap. Here is the paradox at the heart of the report. Even though almost no one fully trusts automated evaluation (Finding 2), two-thirds of organizations (66%) either already allow zero-human-in-the-loop deployment for low-risk agents (34%) or are actively engineering their pipelines to permit it within a year (33%). Only 22% rule it out for the foreseeable future. The direction is unambiguous: enterprises are moving to let evaluations gate production autonomously — removing the human check — at the same moment they say those evaluations don’t reliably match reality. The autonomy ceiling is rising faster than the assurance beneath it, which is the mechanism by which the false-confidence failures of Finding 1 will scale rather than shrink. Notably, the autonomy bet is not just a small company phenomenon. Splitting the sample by company size, larger enterprises are slightly further down the path toward zero human review than smaller companies (70% versus 64%) and slightly more likely to have shipped an evaluation-passing agent that then failed a customer (54% versus 48%). The assumption that large, regulated organizations are holding the human in the loop longest is, in this sample, backwards. To be sure, these are directional figures, since the survey was not a huge sample — 57 respondents from companies with 2,500+ employees and 100 from companies smaller than that. Finding 4: The evaluation stack is fragmented and provider-led Provider-native evals lead — tied with no dedicated tool at all We asked which agent reliability or evaluation platform enterprises primarily use today. The market has no clear leader — and a large share has nothing dedicated. The evaluation layer is early and unconsolidated. Provider-native tooling leads — OpenAI’s native evals and traces (17%) and Anthropic’s Claude Console evals (13%) together outweigh any independent platform — but it is tied at the top by a striking answer: 17% of enterprises use no dedicated agent-evaluation tooling at all, a notable gap for organizations shipping agents to customers. The specialist evaluation vendors — DeepEval (12%), Braintrust (8%), LangSmith, Weave, Promptfoo, Langfuse, Arize — are scattered across single to low double digits, and 11% have built their own. No independent platform has yet become the category standard, which leaves most enterprises evaluating agents with provider-native tools, home-grown scripts, or nothing. Finding 5: Production monitoring rarely watches output quality Only a quarter run real-time quality checks on live traffic Production monitoring for an AI agent can watch two very different things. It can watch whether the system is functioning — is the agent up and responding, did each request complete, how fast, at what cost, with any errors. Or it can watch whether the agent's output is correct — automated checks that evaluate the content of each answer as it goes out: did the agent give the right answer, take the right action, stay within policy. The distinction matters because a confidently wrong answer is invisible to the first kind of monitoring: the request completes, the response is fast, no error is thrown, and every functioning-metric reads healthy. We asked organizations which kind their live production monitoring is built for today. Grouped by what is actually being watched, the split is stark: 51% of organizations monitor only whether the agent is functioning, while 23% monitor whether its answers are right. Counting the ad-hoc reviewers and the don't-knows, roughly three-quarters of organizations run no automated, real-time evaluation of output correctness in production — they can see that the system is up and what it costs, and they are taking the correctness of its answers on faith. That blind spot is the runtime counterpart to the pre-deployment gap in Finding 1: the same organizations engineering the human out of the deployment decision mostly cannot see, in real time, when the deployed agent starts getting things wrong. Finding 6: Bought on cost, measured on consistency Price and integration drive selection; evaluation consistency is the goal We asked what most influenced enterprises’ choice of an evaluation vendor, and what they treat as their primary measure of success. Both answers are pragmatic. Enterprises buy evaluation tooling on economics and trust it on repeatability. Cost of evaluations (28%) narrowly leads selection, just ahead of ease of integration (27%) and evaluation accuracy (24%) — breadth of observability (13%) and vendor roadmap (4%) matter far less. On what success looks like, more than a third (36%) name evaluation consistency — getting the same verdict on the same behavior every time — well ahead of speed of experimentation (19%), reduction in failures (18%), production visibility (13%), and compliance (11%). The emphasis on consistency is telling: before enterprises can trust an evaluation’s verdict, they need it to be stable — the very property whose absence (bias and inconsistency) ranked among the top trust limitations in Finding 2. Satisfaction with current tooling is only moderate, averaging 3.8 on a five-point scale across overall satisfaction, ease of implementation, and value for money. Finding 7: The next dollar goes to humans and observability Investment is flowing to oversight, not just automation We asked which reliability and evaluation investment will grow most over the next year. The money is going toward watching agents more closely — including with people. The second-largest planned investment — behind only production observability — is human review workflows, at 26%. Read against Finding 1, that is the report's quietest contradiction: at the same moment two-thirds of enterprises are engineering the human out of the deployment decision, more of them plan to grow spending on human reviewers (26%) than on the automated evaluation pipelines (16%) that would replace them. The zero-human trajectory and the human-review budget are rising in the same companies at the same time. Indeed, only 8% report that their budget is not increasing. Taken together, enterprises are hedging: building toward autonomy while spending to watch agents more closely and keep humans available for the calls that automated evaluation cannot yet be trusted to make. Finding 8: A tooling reshuffle is coming Nearly two-thirds plan to adopt or switch platforms within a year We asked whether enterprises plan to adopt a new, additional, or replacement evaluation platform, and which they are considering. Few intend to stand pat. The evaluation market is wide open. While 36% have no plans to change, a clear majority (64%) intend to adopt a new, additional, or replacement platform within twelve months, and 31% within the next quarter. The consideration set points where current usage is thinnest: Confident AI’s DeepEval leads what enterprises are evaluating (20%), ahead of OpenAI’s native evals (13%) and Braintrust (9%) — the open-source specialists drawing more interest than their present footprint. Given that so many enterprises today rely on provider-native tools or nothing at all (Finding 4), this is less a defection than a first real wave of tooling adoption — the moment the evaluation layer starts to consolidate. Which platforms earn that trust, in a market where almost no one trusts automated evaluation yet, is the open question this series will keep tracking. The bottom line: An evaluation gap that autonomy will widen, not close Organizations with 100 or more employees are granting AI agents more independence than they trust their evaluations to support. Half have already shipped an agent that passed its evals and then failed a customer; almost none fully trust automated evaluation, chiefly because it doesn’t match real-world outcomes; and most watch production for uptime and cost rather than for whether the agent’s answers are right. Yet two-thirds already allow, or are actively building toward, deploying to production on automated evaluation alone. The vendor market is early and unsettled: the most common primary evaluation tools are provider-native evals, tied with no dedicated tooling at all, and a clear majority plan to adopt or switch platforms within the year. Encouragingly, the next dollar is going to observability and — pointedly — human review, suggesting enterprises sense the gap even as they engineer past it. At 157 respondents in a single wave this is a directional read, skewed toward the mid-market — but the direction is clear: autonomy is being granted on the strength of evaluations that the people granting it do not yet trust. The evaluation gap is not a coverage problem that more tests alone will close; it is a problem of evaluations that reflect reality and can be trusted to gate it. The open question for later waves is whether assurance catches up to autonomy — or whether the false-confidence failures move from customer incidents into changes that deploy themselves. Based on survey responses from 157 qualified enterprise respondents (100+ employees), drawn from a single June 2026 wave. This is a directional read rather than a precise measurement — the sample is self-selected, not a probability sample, and skews toward the mid-market. Respondents include product and program managers, consultants and advisors, directors of engineering/IT, and CIOs/CTOs/CISOs, among other functions, across technology/software, retail/consumer, healthcare/life sciences, manufacturing, and other industries.

ReadSource

arXiv AI/ML / 5:59 PM

arXiv paper: SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators

A new arXiv AI paper by Yuncong Yang, Zhengtao Han, and Furkan Ozyurt, and 6 more studies SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators.

ReadSource

Simon Willison LLMs / 8:19 PM

Kimi K3, and what we can still learn from the pelican benchmark

Chinese AI lab Moonshot AI announced Kimi K3 this morning, describing it as their "most capable model to date, with 2.8 trillion parameters". It's currently available via their website and API, but an open weight release is promised "by July 27, 2026". Moonshot are calling this the first "open 3T-class model" (I guess they're rounding 2.8 trillion up to 3 trillion), taking the crown from DeepSeek's 1.6T v4 Pro . Their self-reported benchmarks have K3 mostly beating Claude Opus 4.8 max and GPT-5.5 high, while losing out to Claude Fable 5 and GPT-5.6 Sol. A few highlights from the Artificial Analysis report on the model: "On our private long-horizon knowledge work evaluation, Kimi K3 reaches an overall Elo of 1547, +732 points from Kimi K2.6 and behind only Claude Fable 5." "Cost per task ($0.94) is similar to GPT-5.6 Sol ($1.04), ~1/2 the price of Opus 4.8 ($1.80) and higher than open weights peers" "Kimi K3’s token usage on the Artificial Analysis Intelligence Index decreased significantly, using 21% fewer output tokens than K2.6." The model is also now the leading model on Arena.ai's Frontend Code arena , surpassing even Claude Fable 5. The new model is notable for the pricing: $3/million input tokens and $15/million output tokens, putting it at the same level as Anthropic's Claude Sonnet series and making it the most expensive model released by a Chinese AI lab to date. This is a significant increase on their earlier models such as Kimi K2.6 at $0.95/$4. 2.8 trillion parameters is also more than twice the size of that 1T model. But how does it pelican? I used OpenRouter (to avoid signing up for a Moonshot API key) with the llm-openrouter plugin to generate an SVG of a pelican riding a bicycle: llm -m openrouter/moonshotai/kimi-k3 'Generate an SVG of a pelican riding a bicycle' Here's the transcript . It looks like this: That pelican took 95 input tokens and 16,658 output tokens (13,241 were reasoning tokens), for a total cost of 25 cents ! Since K3 accepts image input I ran it against that rendered SVG above (with my alt text prompt ) and got back (for 0.6 cents ): Cartoon illustration of a white pelican wearing a red scarf, riding a red bicycle along a gray road with white dashed lines; the pelican has a large orange beak and webbed orange feet pedaling, with white motion lines behind it; the background shows a light blue sky with white clouds, a yellow sun, two small black birds in flight, and green grass with tiny white flowers in the foreground What can we learn from the pelican? My Generate an SVG of a pelican riding a bicycle test is 21 months old now. It was never a particularly great benchmark. It started out as a joke on how absurdly difficult it is to compare these models, but then for the first year it turned out to have a surprising correlation to how good the models actually were. That connection has been mostly severed now. The GPT-5.6 and Claude Fable 5 pelicans are outclassed by GLM-5.2 , and much as I love GLM I don't think that's a Fable-class model. (I'm still not convinced that labs are training for the benchmark - if they were, I'd expect much better results. There's a chance that Gemini has optimized for any combination of an animal on a vehicle though!) The biggest limitation of the pelican is that it doesn't touch at all on the thing that matters most for today's model: agentic tool calling and the ability to operate tools reliably as conversations grow in length. So don't go using pelicans to compare models! All of that said, I still get a decent amount of value out of running the benchmark myself. Firstly, it's a forcing function for actually trying the model. If I show you a pelican, that means I've managed to run a prompt through it. If the model has an official API I'll use that, if it's open weight (and small enough to fit a 128GB M5 MacBook Pro) I'll try running it on my own machine, usually via llama.cpp or LM Studio or Ollama . I'll frequently use OpenRouter since that usually provides a proxy to an official API without me needing a new API key. Most of my pelicans are generated using my LLM CLI tool , which helps encourage me to ensure the latest models are supported by that (via one of its plugins). More importantly though, even the act of a single prompt to "Generate an SVG of a pelican riding a bicycle" can reveal interesting model characteristics. Consider the result for Kimi K3 today. Running those simple prompts helped emphasize several points about the model. It only has one reasoning effort right now, "max" - and it shows. The model consumed 13,241 reasoning tokens to output 3,417 tokens of response. This is expensive - the pelican cost 25 cents! How does the prompt "Generate an SVG of a pelican riding a bicycle" add up to 95 input tokens? OpenAI's tokenizer counts 10, Anthropic's counts 10 for Opus 4.6, 30 for Opus 4.7 and 25 for Sonnet 5/Fable 5. Prompting "hi" to Kimi K3 counted 86 tokens, suggesting there may be an 85 token hidden system prompt. It refused to leak it though. Vision works well: the alt text it generated is very good. K3 currently only has one thinking effort level, but I've been deriving quite a bit of value recently from running the same pelican prompt through different effort levels to get a quick idea for what impact those have. Here's my matrix for the GPT-5.6 model family , for example. Really though the main things I gain from the pelican test are: It's a "hello world" exercise for prompting a model A rough cost and reasoning estimate for a simple task Confirmation that the model can output valid SVG and has a basic idea of geometry and spatial awareness. This is a much bigger deal for the smaller models that run on my laptop. It's still interesting to compare pelicans between releases in the same model family. K3's pelican is a notable improvement from Kimi 2.5 . It's something I can share that demonstrates I've tried it. Plus a comment with a pelican in it is kind of a tradition on Hacker News at this point, any time I'm late I get comments asking where it is! Tags: ai , generative-ai , llms , llm-pricing , pelican-riding-a-bicycle , llm-release , ai-in-china , artificial-analysis , moonshot , kimi

ReadSource

arXiv AI/ML / 5:50 PM

arXiv paper: Mask Forcing: Improving Autoregressive Video Diffusion Distillation via Dual-Noise Masking Rollout

A new arXiv AI paper by Zhuoran Zhao, Shengju Qian, and Tongtong Liang, and 7 more studies Mask Forcing: Improving Autoregressive Video Diffusion Distillation via Dual-Noise Masking Rollout.

ReadSource

arXiv AI/ML / 5:27 PM

arXiv paper: What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation Policies

A new arXiv AI paper by Vivek Chavan, Pengtao Xie, and Yahuan Shi, and 3 more studies What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation Policies.

ReadSource

arXiv AI/ML / 5:58 PM

arXiv paper: Robust PAC Learning of Concurrent Stochastic Games

A new arXiv AI paper by Angel Y. He and David Parker studies Robust PAC Learning of Concurrent Stochastic Games.

ReadSource

arXiv AI/ML / 5:57 PM

arXiv paper: Mechanism Design for Alignment and Control

A new arXiv AI paper by Dirk Bergemann, Andrew Koh, and Stephen Morris studies Mechanism Design for Alignment and Control.

ReadSource

arXiv AI/ML / 5:39 PM

arXiv paper: Scaling Near-Optimal SFT-RL Annotation Budget Allocation from Small to Large LLMs

A new arXiv AI paper by Jingtan Wang, Arun Verma, and Xiaoqiang Lin, and 4 more studies Scaling Near-Optimal SFT-RL Annotation Budget Allocation from Small to Large LLMs.

ReadSource

arXiv AI/ML / 5:59 PM

arXiv paper: Constant Individual Regret in General Games

A new arXiv AI paper by Mingyang Liu, Gabriele Farina, and Asuman Ozdaglar studies Constant Individual Regret in General Games.

ReadSource

arXiv AI/ML / 11:18 AM

arXiv paper: Generative Retrieval for E-commerce: Jointly Learning Embedding and Codebook with Same Product Cluster

A new arXiv AI paper by Songtao Fang, Zihao Xu, and Shaowei Wei, and 2 more studies Generative Retrieval for E-commerce: Jointly Learning Embedding and Codebook with Same Product Cluster.

ReadSource

arXiv AI/ML / 4:54 PM

arXiv paper: Learning the Target Priors Before Image Translation: A Decoupled Training Paradigm for Cross-Modal Image Translation in Remote Sensing

A new arXiv AI paper by Keyan Hu, Mingtao Wang, and Ziyu Zhou, and 4 more studies Learning the Target Priors Before Image Translation: A Decoupled Training Paradigm for Cross-Modal Image Translation in Remote Sensing.

ReadSource

arXiv AI/ML / 5:18 PM

arXiv paper: Token-Level Advertising

A new arXiv AI paper by Hanbing Liu, Bowei Zhang, and Changyuan Yu, and 2 more studies Token-Level Advertising.

ReadSource

arXiv AI/ML / 5:54 PM

arXiv paper: Truthful Calibration Measures for Sequential Prediction

A new arXiv AI paper by Anagha Gokul, Jason Hartline, and Lunjia Hu, and 2 more studies Truthful Calibration Measures for Sequential Prediction.

ReadSource

arXiv AI/ML / 5:28 PM

arXiv paper: Unified Branch-and-Bound Search for the Steiner Traveling Salesman Problem on Graphs of Convex Sets

A new arXiv AI paper by Jingtao Tang and Hang Ma studies Unified Branch-and-Bound Search for the Steiner Traveling Salesman Problem on Graphs of Convex Sets.

ReadSource

arXiv AI/ML / 4:38 PM

arXiv paper: Detecting Backdoors in Object Detection via Pre-NMS Prediction Distribution Shift

A new arXiv AI paper by Longtian Wang, Zhengyu Zhao, and Chenhao Lin, and 5 more studies Detecting Backdoors in Object Detection via Pre-NMS Prediction Distribution Shift.

ReadSource

arXiv AI/ML / 5:53 PM

arXiv paper: EDITBRIDGE: Towards Faithful and Efficient Ultra-High-Resolution Image Editing

A new arXiv AI paper by Jiayi Song, Shijie Huang, and Fangtai Wu, and 5 more studies EDITBRIDGE: Towards Faithful and Efficient Ultra-High-Resolution Image Editing.

ReadSource

arXiv AI/ML / 5:52 PM

arXiv paper: The concentration game: Bayesian updating, regret, and information

A new arXiv AI paper by Akshay Balsubramani studies The concentration game: Bayesian updating, regret, and information.

ReadSource

arXiv AI/ML / 5:57 PM

arXiv paper: AV-AIVAT: 74x Cheaper Agent Evaluation with Certified Anytime-Valid Stopping in Imperfect-Information Games

A new arXiv AI paper by Boning Li, Yu Chen, and Longbo Huang studies AV-AIVAT: 74x Cheaper Agent Evaluation with Certified Anytime-Valid Stopping in Imperfect-Information Games.

ReadSource

arXiv AI/ML / 5:53 PM

arXiv paper: Resourced Authority A Mechanism-Design Model for Participatory Governance of Deployed AI Agents

A new arXiv AI paper by Praphul Chandra, Sujit Gujar, and Ganesh Ghalme studies Resourced Authority A Mechanism-Design Model for Participatory Governance of Deployed AI Agents.

ReadSource

arXiv AI/ML / 5:37 PM

arXiv paper: Algorithms for Structured Elections under Thiele Voting Rules

A new arXiv AI paper by Alexandra Lassota and Krzysztof Sornat studies Algorithms for Structured Elections under Thiele Voting Rules.

ReadSource

Latest story in this edition: 3:01 PM

Back to front page