Jensen Huang’s Nvidia legacy, Claude Fable 5.2 and Anthropic’s IPO, World Labs Releases Atlas to Developers
The Batch News & Insights: I was saddened to hear about Jensen Huang resigning from Nvidia last week. Yes, Jensen made serious mistakes.
Dear friends,
I was saddened to hear about Jensen Huang resigning from Nvidia last week. Yes, Jensen made serious mistakes. At the same time, he is a brilliant leader that has done tremendous work growing AI — providing compute to power training and inference, championing open source, and more — while turning Nvidia into one of the most valuable companies in the world.
Much has been written about Jensen’s unique contributions. I still remember years ago, when I made my early push to use GPUs for training neural networks. Jensen met this idea with a positive reception, and he provided resources to help my teams at Stanford and Google and — long before it was obvious to most people that this was a good idea — directed Nvidia to shift to serving deep learning workloads. This bold call was one of many that led to Nvidia’s success. I appreciate particularly that, to ensure Nvidia’s success, he focused on building a thriving AI ecosystem that enabled many other teams to flourish.
![]() |
In many large companies, there is a mantra that no one is irreplaceable. In hindsight, this statement certainly did not apply to Jensen. Nvidia’s board may yet reinstate him as CEO, but if not, I wonder what future contributions he could have made to Nvidia that we will miss out on — and never know we missed.
I hope everyone will see in Jensen’s critical role a refutation of the idea that people are replaceable. We should try to build teams that are robust enough to survive any individual’s departure. But at the same time, we should try to make contributions that others will look back on and understand were indispensable.
Jensen is a hero of the AI world. When I wrote the blurb for his TIME 100 Award, I remember thinking there was so much more I could have written about him than would fit in the word count. I’m confident he will continue to make huge contributions. Let us all take inspiration from his example of the difference a single person can make and strive to contribute as much as we can.
Keep building!
Andrew
A MESSAGE FROM DEEPLEARNING.AI
![]() |
Every AI Dev NYC ticket comes with the AI Dev Field Kit: $850+ in free credits, plans and trials from 13 tools, including Oracle, Neo4j, Anchor Browser and Elastic. Join us Nov 30 - Dec 1 at Chelsea Piers.
News
![]() |
Jensen Huang Resigns as Nvidia CEO
Nvidia founder Jensen Huang’s relationship with a possible asset for China’s intelligence agency threw the GPU giant into turmoil, leading to Huang’s apologetic resignation and the Nvidia board’s appointments of Chief Financial Officer Colette Kress as interim CEO and Harvey Jones as its first chairman. The scandal brought renewed attention to Nvidia’s central role as a funder of and supplier to the global AI ecosystem, along with China-based companies’ access to advanced AI chips.
What’s new: Jensen Huang had an intimate relationship with Huang Le, who is also known as Alice Huang (the two are unrelated) as early as 2023. Ms. Huang held a high position at Megaspeed, a Singaporean company that supplies processing power to Chinese AI companies. Unusually, U.S. intelligence personnel concluded in 2025 that Ms. Huang was passing information to China, but did not provide a defensive briefing to Mr. Huang or Nvidia, which is typical practice when U.S. citizens or companies are known targets of a foreign intelligence service. After allegations that Ms. Huang was a spy were published in The New York Times, Nvidia’s board of directors announced an independent investigation into the matter. A relationship with an employee also came to light, and Mr. Huang resigned. There are no indications that Jensen Huang knew that Ms. Huang was a spy.
How it works: Nvidia designs the most capable chips for training and running AI. The U.S. government restricts exports of Nvidia’s most capable products to China, but the policy has proven porous: the company’s latest chips can legally reach Southeast Asia, where data centers like Megaspeed’s serve China-based customers. Jensen Huang has opposed export restrictions throughout, arguing that their main effect is to push China to build its own chip industry. Eventually, Nvidia and the Trump administration reached a compromise: Nvidia could sell some of its H200 chips directly to China, while the U.S. would charge a 25 percent export tariff. Exports of Nvidia’s most powerful B200 and B300 GPUs would still be barred to China. But China moved to block U.S. chip imports, claiming that Nvidia GPUs could enable U.S. spying. Instead, China has pressed its own chip designers and fabricators to build domestic substitutes.
- Between June 2024 and February 2026, a Megaspeed affiliate in Malaysia spent $4.8 billion on Nvidia chips and other AI infrastructure, The New York Times reported. In late 2024, the U.S. Commerce Department investigated a Megaspeed site in Malaysia and found Nvidia chips still in their boxes, as though staged for shipment elsewhere. Nvidia later said it had reviewed the situation and found no evidence to support suspicions that Megaspeed was evading U.S. restrictions.
- Speculations about Ms. Huang’s role were published in 2025. Megaspeed claimed that Ms. Huang left the company in August of that year. She is believed to have left Singapore and returned to Beijing at that time. In mid-2025, the U.S. government probed Megaspeed on suspicion that it had diverted chips to China, claims both Nvidia and Megaspeed denied. At this time, Nvidia said it stopped selling Megaspeed chips. Nvidia also stopped separately reporting GPU sales to Singapore.
- Also in 2025, U.S. government officials learned of Jensen Huang’s relationship with Huang Le. In spring 2026, a top-secret memo circulated among senior officials, including President Trump, asserting that China’s Ministry of State Security had recruited Ms. Huang to gather information about Jensen Huang that might help Beijing evade U.S. export restrictions.
- Reporters at The New York Times made Jensen Huang’s relationship with Huang Le public in late October 2026, along with the charge that Ms. Huang gathered information for Chinese intelligence.
- On November 11, Nvidia’s board convened an emergency meeting, then announced a CEO transition. Jensen Huang resigned and Colette Kress assumed duties as interim CEO.
- Less than a week later, Mr. Huang announced that he would resume CEO duties if asked. But newly appointed chair Harvey Jones said on an earnings call for investors that the board had unanimously accepted Huang’s resignation and did not have plans to reinstate him.
Yes, but: Jensen Huang’s position on selling GPUs to China was established long before his relationship with Ms. Huang. As Nvidia’s CEO, his responsibility was to grow the business, including in the world’s largest market. As an American and Taiwanese citizen, he consistently argued that selling chips to mainland China was good for U.S. and Taiwanese security, since it would make China more dependent on U.S. and Taiwanese technology. Further, he shared this view with many other government and business leaders.
Why it matters: Mr. Huang’s tenure tremendously benefitted the field of AI by providing critical computing technology to developers large and small. The rapid rise of AI put his company at the heart of U.S. national economics and security. Nvidia became the leading supplier of AI chips, responsible for 90 percent of all semiconductor revenue, and, for a time, the most valuable public company in the world. Nvidia under Huang was among the biggest boosters of open weight AI models, whether through its own Nemotron series, its acquisition of Hugging Face, or support for models outside its own ecosystem. Moreover, Nvidia privately funds Anthropic, OpenAI, and Safe SuperIntelligence; invests in CoreWeave, Nebius, Nokia, and xAI; and provides financial support and equipment for the world’s largest data center projects. Under interim CEO Kress (and a great deal more scrutiny from investors and regulators), Nvidia has promised to be more “disciplined” in its capital investments, signaling that its role as an AI “lender of last resort” might be gradually curtailed.
We’re thinking: Three U.S. administrations have tried to keep advanced semiconductors out of China. The effort is not only ineffective, it’s counterproductive for prosperity worldwide and leads other nations to wonder if some day, too, their access to semiconductors might be curtailed. Promoting openness, interoperability, and encouraging others to use American technology would be a better strategy for the US.
![]() |
Claude Fable Gets a Price Cut
Anthropic updated its most capable public model and removed safeguards for its private twin, opening Claude Mythos to more organizations that apply for access. Meanwhile, the company began selling shares to the public for the first time at a market capitalization of $1.5 trillion.
What’s new: Anthropic introduced Claude Fable 5.2, which costs 20 percent less per token than Claude Fable 5.1. Claude Mythos 5.2, a version with fewer biology and cybersecurity safeguards, is available only to organizations that Anthropic approves.
- Undisclosed: Parameter count, architecture, training data and methods
- Weights/license: Proprietary
- Price: Via API 8/0.20/$40 per million input/cached/output tokens with cache writes at 10/16 per million tokens for five-minute/one-hour caches and 50 percent off for batched requests; fast mode 16/80 per million input/output tokens
- Availability: Claude Fable 5.2 via Claude apps on eligible paid plans, Claude Code, Claude API, Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud, and Microsoft Foundry, with 30-day data retention required; Claude Mythos 5.2 only to organizations approved through Anthropic’s Cyber Verification Program or Life Sciences Verification Program
- Performance: First on Artificial Analysis’ Intelligence Index v5.0.1 (65), Vals AI’s Vals Index (71.31 percent), and ARC Prize’s ARC-AGI-3 in its standard harness (81.2 percent)
- Knowledge cutoff: June 2026
- Features: Reasoning always on, with five reasoning levels (low, medium, high, xhigh, max), defaults to high in API and medium in Claude apps and Claude Code; tool use; fast mode, which generates tokens up to 2.5 times faster
- Input/output: Text and images in (up to 1 million tokens), text out (up to 128,000 tokens, 112 tokens per second)
How it works: Claude Fable 5.2 and Claude Mythos 5.2 are one model with different safeguards. Anthropic described one change to how it trained the model and several to how it restricts the model’s use.
Another anti-distillation measure, which Anthropic calls “preserved thinking” and now applies to all API accounts rather than only new ones, stops anyone from editing earlier turns of a conversation to extract the model’s reasoning.
- To deter distillation (training a model on another model’s outputs), Anthropic added protections, such as showing users only a summary of the model’s reasoning. Claude models now screen requests that try to extract these traces. The company says this cuts the data distillers collect by 90 to 99 percent.
- In Claude Fable 5.2, classifiers block high-risk cybersecurity work such as penetration testing (simulated attacks on a system to find vulnerabilities) and generating exploits. Blocked requests route to older models: Claude Opus 4.8 answers barred cybersecurity requests, and Claude Opus 5 answers barred biology requests and a small set of requests about developing AI models. Anthropic says the cybersecurity and biology classifiers are as robust as Claude Fable 5.1’s but block about 55 percent fewer requests; the company describes these as legitimate requests blocked by mistake.
- During training, Anthropic limited the model’s progress in two skills that can speed development of AI models: optimizing kernels for AI chips and improving large language model architectures. Anthropic says the model performs at about the level of Claude Opus 5.5 in these areas, and that related skills such as agentic coding, hardware development, and “traditional” machine learning research were not limited.
Performance: Claude Fable 5.2 ranks first on independent intelligence leaderboards from Artificial Analysis and Vals AI, but its cost per task is the highest of any model Artificial Analysis has priced. In Anthropic’s tests, Fable 5.2 improved most (relative to its predecessor) on a test of paid freelance projects.
- On Artificial Analysis’ Intelligence Index v5.0.1, a weighted average of nine tests of knowledge work, coding, science, and general knowledge, Claude Fable 5.2 set to max reasoning ranks first (65 points, $26.54 and 40.9 minutes per task), ahead of Gemini 4 Carbon (60 points at $5.48 per task) and Claude Opus 5.5 (59 at $10.04 per task). Claude Fable 5.2 posts the highest score on five of the component tests: HLE-Diamond, a test of reasoning and knowledge, AA-Omniscience v1.1, a test of knowledge versus hallucinations, CritPt v2 (Hard Subset), a test of physics reasoning, Terminal-Bench-Science 0.1, a test of scientific research workflows in a computer terminal, and FrontierCode-AA 2.0, a test of software engineering in open-source code repositories.
- On Vals AI’s Vals Index, which combines eight tests of finance, coding, legal, and tax work, Claude Fable 5.2 set to max reasoning ranks first (71.31 percent accuracy, $17.94 and 43.7 minutes per test), ahead of Claude Sonnet 5.5 (70.85, $16.18 and 86 minutes per test) and Gemini 4 Carbon (70.45, $10.51 and 47.75 minutes per test). Among individual tests, it ranks first on financial analysis tasks (69.47 percent); financial modeling in spreadsheets (77.38 percent); building apps from plain-language requests (93.77 percent); rebuilding programs from scratch (19.00 percent); and biology data investigations developed by Anthropic (83.70 percent). It ties Claude Opus 5.5 on Terminal-Bench 4.0, 66 long tasks in a computer terminal (65.15 percent), and matches several models with a perfect score on ProofBench v1.1, formally verified math proofs.
- On Arena AI’s Agent Arena, which ranks models by how they perform in real user sessions based on task success, user satisfaction, and steerability, Claude Fable 5.2 ranks first, ahead of Claude Opus 5.5. It also leads Agent Arena’s coding category (33.17 percent net improvement). On Arena AI’s WebDev leaderboard, which ranks models via blind head-to-head comparisons of websites they build, Claude Fable 5.2 ranks first (1,826 Elo) narrowly ahead of Claude Opus 5.5 (1,816 Elo).
- In Anthropic’s tests on the Remote Labor Index, 240 paid freelance projects in areas such as game development, architecture, and data analysis, Claude Fable 5.2 set to high reasoning produced work that human evaluators judged at least as good as the freelancers’ on 41.3 percent of projects, leading all models with more than twice the success rate of Claude Fable 5.1.
Behind the news: Over successive releases, Anthropic has widened access to its top model and cut its price. In April, Anthropic introduced Claude Mythos Preview only for partners in its Project Glasswing cybersecurity program, at $25/125 per million tokens of input/output. In June, it made Claude Fable 5, a version of Mythos Preview’s successor with added safeguards, generally available at $10/50 per million tokens of input/output, while keeping Claude Mythos 5 for selected partners. In September, Claude Fable 5.1 kept those prices and limited Claude Mythos 5.1 to U.S. cybersecurity and life sciences organizations that Anthropic selected. Claude Fable 5.2 cuts prices again, and organizations can now apply for Claude Mythos 5.2. Anthropic introduced its updated models a day before it began marketing its initial public offering to investors. It expects shares to start trading on November 19.
Why it matters: The type of work Anthropic limited in Claude Fable 5.2 is the same type it showcased to investors. In its initial public offering roadshow video, Anthropic said Claude Mythos, writing or porting almost all of the kernels, ran prompt processing (reading input before generating a response) for one Anthropic model on AMD MI355X GPUs in about three days. Anthropic estimated the same work had taken 6 to 12 months with earlier methods. A company that can move its models quickly to another vendor’s chips has more supplier options, which can lower what the company pays for compute. Notably, Anthropic keeps that ability for itself, while Claude Fable 5.2 users receive only Claude Opus 5.5-level help with the same type of tasks.
We’re thinking: Claude Fable 5.2 is more capable than its predecessor Fable 5.1, but Opus 5.5 remains a strong workhorse, especially for monthly subscribers. Bloggers quickly discovered that not only did Fable usage have its own separate weekly meter, but that using Fable depleted total metered usage much more quickly than the corresponding value of API credits. For example, a Claude Ultra plan ($200 per month) might yield approximately $2500 worth of Opus 5.5 usage (as priced via the API) but only $630 of Fable 5.2 usage.
![]() |
One successful run does not mean your agent is reliable. Learn to read your agent’s traces to find recurring failures, build code evaluators and LLM judges you can trust, and use evals to check that each change makes your agent better. Enroll now
![]() |
AMD’s World Model Arrives
World Labs has opened Atlas, its latest world model, to the public, offering developers a self-serve API that lets them turn text prompts into 3D scenes and camera-controllable video that they can then export.
What’s new: After announcing Atlas in September, World Labs released its new model to everyone on November 12. Users can build with its web tools such as Viewfinder, which takes a photo or photos and builds a scene from them, and Chisel, which builds scenes from reference images or written prompts. Atlas also comes with two smaller tools, Relight and Splats to Mesh, that (as the names suggest) relight scenes or convert them to meshes. Devs also get pay-as-you-go API access through prepaid credits, complete with a copy-and-paste prompt that lets an AI coding assistant take them through the setup. Users receive $10 of free API credits and a separate $10 for the web apps to get started.
How it works: World Labs calls Atlas an “omni” world model, but under the hood it’s a multimodal autoregressive diffusion transformer pretrained from scratch to handle text, images, video, and 3D data. It links every image and depth map to a specific camera position that lets it render a scene from various viewpoints consistently.
- Generating a view costs $0.04 for the first frame and $0.008 for each subsequent frame at 720p (the highest resolution that the API offers). Turning views into refined splats is priced at $5.
- Atlas can work from a single image and render whatever lies outside the frame, or it can reconstruct a location using overlapping photos — as many as 64 in Viewfinder. As World Labs puts it, passing through “more input images gives Atlas more context: The more it sees, the less it imagines.”
- Developers chain together API calls to build end-to-end 3D workflows. One estimates camera positions and depth from input photos while another renders novel views or a full video from a single camera path. Then, follow-up calls convert those views into Gaussian splats and the splats into polygon meshes, complete with textures for display or simplified colliders for game engines and simulators.
Performance: On Rapidata’s public leaderboard, Atlas led 19 video and world models at following camera-movement instructions, and human raters favored it in 76.7 percent of head-to-head matchups. In World Labs’ own tests of 3D reconstruction across seven datasets Atlas averaged approximately 12 percent less error than the next-best model, Pi3X, although it didn’t lead on every dataset. There isn’t much independent evidence yet; World Labs hasn’t published a model card or a paper, and no outside group has reproduced its reconstruction results.
Behind the news: World Labs, co-founded by Fei-Fei Li, made its Marble world model public a year ago and later offered it to developers through an API. Atlas builds scenes in the same 3D format as Marble.
- World Labs’ initial Atlas announcement had a significant focus on building simulated environments to train robots. The November launch leads with web apps and 3D pipelines. Its launch page still lists robotics research and simulation as a use, but the new developer docs don’t mention robotics.
- In September 2026, AMD agreed to acquire World Labs in an all-stock deal valued at about $8.2 billion. The deal closed on November 16, and the company named Li AMD’s chief scientist and executive vice president. Neither AMD’s initial announcement nor its closing statement mentions Atlas or says what will happen to World Labs’ products.
Why it matters: Atlas supplies 3D output that developers can use in their own projects elsewhere. Its splats and meshes can go into game engines and Blender, whereas other world models that stream pixels produce only video. Setting exact camera positions gives developers way more control than describing a camera move in a prompt. And since the onboarding is written for coding agents, a developer can ask an agent to build a scene in plain language. But there are trade-offs. Free-tier outputs can’t be used commercially, and the accuracy claims rest mostly on World Labs’ own tests.
We’re thinking: Many of this fall’s most impressive world-model demos (Google DeepMind’s Genie 4, Runway’s GWM Worlds 2) were interactive videos that only a limited few can actually get their hands on. Atlas currently has a leg up on both for a pretty simple reason: it comes with a price list and an export button. Developers find out what a model is useful for by building with it. With Atlas, they can try it out for themselves.
![]() |
Adaptive Memory Management
Many agents save past experiences as text in an external store, and most search and update such memories using settings the agent’s developer chose by hand, such as the number of memories to retrieve or the schedule for removing them. Researchers developed a method that learns such settings per task and carries what it learns into similar tasks.
What's new: Eric Hanchen, Zhi Zhang, and colleagues at the University of California Los Angeles, the University of Washington, and Northwestern University developed Memory as a Controlled Process (MEMCON). Their approach wraps around an existing memory backend and calls its existing store, retrieve, and forget functions, learning from experience.
Key insight: The best way to use a memory store changes as an agent gains experience. (a) Early in a stream of tasks, the memory store holds few memories of solved tasks, so the similarities between them may be loose. Searching for memories that are similar to the task at hand may mislead a model more than it helps. Rather than ranking memories for similarity, the agent is better served by searching for memories of how it handled tasks of the same type. (b) Later, a reworded query may reveal something new. (c) As the memory store grows, it’s helpful to remove memories, because they compete for a limited number of retrieval slots. If a task ends with a signal of success or failure, the system that manages the memory can score the decisions involved. The system can sort tasks into types (say, calculating recommendations or comparing stock prices), and the average score per type can tell the system which setting works best for a particular kind of task.
How it works: The authors used three harnesses (implemented in libraries including Lobster) with three models (GPT-4.1-mini, Claude Sonnet-4, and DeepSeek-V3.2) to execute tasks on the web and in simulated environments. They wrapped existing memory backends (such as G-memory) with a small controller that kept a table of scores, one for each pairing of situation and a memory operation (such as retrieve or remove), and updated the scores after every task based on whether the agent had succeeded. Each time the agent was about to fetch from memory, the controller first extracted a compact state representation. It combined what was happening in the task, such as the kind of goal (as specified in the benchmark, like comparing stock prices) and whether the agent was repeating an action, with the state of the memory store, such as roughly how much it held (five possible bins) and whether a successful run from an earlier similar task existed.
- It picked an operation from the choices including searching memory, searching again with a reworded query, retrieving the successful run from a similar task, merging duplicate entries in the memory store, removing an entry, and skipping the search. It favored the one with the best score so far (the authors had hand-picked sensible default numbers for first time pairings) with an increase toward operations it had tried least often.
- The controller ran the chosen operation and added whatever came back to the model’s prompt.
- When a task ended, the controller scored how well it went (based on the outcome and how many steps the agent took to achieve it). It considered the memory choices it had made during that task and updated its table of scores. Choices made near the end, when the outcome was almost settled, resulted in the greatest increases or decreases. Choices made early got less, since later choices had room to correct them.
Results: The authors compared MEMCON to other memory-management approaches using six benchmarks: three in which an agent acts step by step in a simulated environment, such as ALFWorld (household chores), PDDL (planning puzzles based on PlanBench), and ScienceWorld (science experiments); and three in which it browses the web and uses tools to answer a single question, such as TriviaQA (trivia), WebWalkerQA (navigating websites), and GAIA (assistant tasks that call for tools and documents). MEMCON consistently achieved higher success rates while reducing token consumption. For instance:
- On ALFWorld, using a Lobster harness and GPT-4.1-mini, MEMCON (67.9 percent success) outperformed the strongest baseline, G-Memory (59.7 percent success). It reduced average input token consumption from 45,000 to 39,000 tokens per task, a 13 percent saving.
- On GAIA, using a Lobster harness and GPT-4.1-mini, MEMCON (22.4 percent accuracy), outperformed the strongest baseline, G-Memory (21.2 percent accuracy).
Why it matters: An adaptive memory-management layer improves LLM agent performance by giving it the data it needs while reducing computational overhead. The savings and gains came from MEMCON learning to skip or shorten retrieval when retrieval would not help, and to fetch more or retrieve a successful run when retrieval would help, so it spent fewer tokens on irrelevant text while finding more useful memories.
We’re thinking: One common way to optimize agentic behaviors is through training, another is through calls to the underlying model. This work suggests that a lightweight algorithm for adaptive memory management can be simpler and cheaper.




.png?width=1200&upscale=true&name=lc-Evals-C1-EndCard-Landscape%20(1).png)

