AI Daily · November 13 Friday

2026-11-12 08:00 — 2026-11-13 08:00 (UTC+8)

1. 01

   WSJ: Nvidia's board is preparing to oust Jensen Huang; shares fall about 8% after hours

   Nvidia's board is preparing to oust co-founder and CEO Jensen Huang and could meet as soon as Wednesday to decide his future, the Wall Street Journal reported, citing people familiar with the matter. Huang earlier admitted to infidelity, and the board is investigating whether he was targeted by a foreign intelligence operation. Shares fell about 8% after hours.

   Read the full report

   The founder who built the world's dominant AI chip supplier may be pushed out. According to Reuters, citing a Wall Street Journal report on Wednesday, November 11 (US Eastern time), Nvidia's board is preparing to oust co-founder and CEO Jensen Huang and could meet as soon as that day to decide his future. The Washington Post, citing two people familiar with the matter, also reported the board would meet Wednesday evening to weigh his fate.

   The saga began on Sunday, when Huang told employees in a memo that he had been unfaithful to his wife and took full responsibility, without giving details. According to the New York Times and other outlets, the woman was a former executive at Singapore-based data center company Megaspeed whom US intelligence links to Chinese intelligence and to facilitating shipments of Nvidia chips to Chinese firms; Beijing called the espionage allegation "groundless speculation". The same day, Nvidia's board opened an investigation, overseen by independent directors with outside counsel, into whether Huang had been targeted by a foreign intelligence operation, while still expressing confidence in him. On Wednesday, the New York Times reported a second affair, with a company executive who reported to Huang, and talk of an ouster escalated.

   Markets reacted sharply: Nvidia shares fell about 8% after hours on Wednesday, and Reuters said the company had lost about $600 billion in market value this week. TechCrunch, citing Bloomberg, said Nvidia is working with an outside firm to identify potential replacements and is considering an interim CEO. Analysts note Huang runs a highly centralized organization with about 60 direct reports and the company has no public succession plan.

   The ouster plan comes from media reports citing people familiar with the matter; Reuters said it could not immediately verify them, and Nvidia and the people involved did not immediately respond to requests for comment. As of the reports this story relies on, the board had not announced a final decision. Nvidia reports earnings on November 17.

2. 02

   Baidu launches Ciyuansuo, a token exchange that verifies and credit-rates model resellers

   At Baidu World 2026, Robin Li said AI's next act is the "intelligent organization" and launched Ciyuansuo, a marketplace for model tokens that verifies suppliers via model fingerprinting and capability and availability tests, then assigns credit ratings. Baidu says it handles over 50 million calls and 700 billion tokens a day after one month.

   Read the full report

   The same model can come from resellers at different prices, with unclear provenance and uneven quality. Baidu wants an exchange to fix that. At Baidu World 2026 in Beijing on November 12, Baidu officially launched Ciyuansuo (roughly "token exchange"), a platform where businesses and individuals can find, buy and use verified model tokens.

   According to Xinhua, Robin Li said in his keynote that new models were released almost daily in the first nine months of the year, but token resellers "take no responsibility for model quality or service stability", making choices harder as models multiply. Ciyuansuo focuses on verification and credit: it checks resources with model fingerprinting, capability tests and performance and availability tests, then uses periodic spot checks and real-time monitoring of calls to give model providers credit ratings. Baidu says that one month after launch, more than 100 supply resources have joined, with over 50 million calls and more than 700 billion tokens per day.

   The keynote's theme was the "intelligent organization". Li compared AI to new electricity and argued it has not delivered expected economic gains because companies "only replaced the motor without redesigning the factory". He proposed three principles of organizational evolution: break down job boundaries, capture tacit knowledge and design closed-loop processes, predicting jobs will converge into Builders who create and Sellers who connect with people.

   Baidu also released the ERNIE Lightning deep-thinking model, the Yijing micro-expression model, a Tianchi 1,024-card supernode for Baidu AI Cloud, and a short animated drama studio for its digital-human agent Yijing (which Baidu says costs about 0.2 yuan per second using its own video model). Baidu said its Apollo Go robotaxi service now covers 28 cities with over 350 million autonomous kilometers. All figures are Baidu's own.

3. 03

   Kunlunxin roadmap: M100 is out, flagship M300 arrives next year, new N architecture in 2028

   At Baidu World 2026, Baidu's chip unit Kunlunxin said its cost-focused M100 has launched as planned, the flagship M300 will ship in 2027, and several chips on a new N architecture will arrive in 2028. Robin Li said scaling AI agents is driving an explosion in inference demand, which will be the main long-term driver of AI chip growth.

   Read the full report

   Baidu's chip company Kunlunxin laid out its product cadence for the next three years at Baidu World 2026 on November 12, built on the view that AI chip growth is shifting from training to inference.

   According to on-site reports from IThome and the Beijing News, Robin Li said AI compute demand used to be dominated by training but is now split between training and inference, and the scaling of agent applications has triggered an explosion in inference demand. "We believe inference will become the primary driver of sustained, long-term growth in AI chips."

   The plan has three steps: the M100 has launched as planned, focused on cost-effectiveness; the flagship M300 will ship next year, with improvements Kunlunxin describes across compute, memory and interconnect; and by 2028 several chips based on a new N architecture will reach the market for different application scenarios. Baidu's official X account gave the same timeline.

   Li called Kunlunxin the result of more than a decade of sustained investment and said its commercial value is increasingly recognized. Public information does not include M300 specs, process node or pricing, and its gains are described only qualitatively, so its competitiveness will depend on published specs and customer deployments.

4. 04

   UK AI chip startup Fractile raises $720M to stack DRAM on chips for faster inference

   British AI chip startup Fractile raised a $720M Series C led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with the UK's Sovereign AI fund participating. The UK government put the valuation above £5 billion; FT and Reuters sources said about $7 billion. Fractile plans to ship its first chips and systems next year.

   Read the full report

   Another inference-focused AI chip company has landed a huge round. British startup Fractile announced on November 11 a $720 million Series C led by Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint. Britain's Sovereign AI fund and the British Business Bank also took part, with the UK government saying the backing aims to ensure the country has a greater stake in the hardware underpinning its future economy.

   The government put the valuation at "over £5 billion"; according to TechCrunch, figures from the Financial Times and a Reuters source put it at about $7 billion. TechCrunch says this makes Fractile one of the best-funded purpose-built AI chip startups alongside Axelera and Etched.

   Fractile's pitch is its hardware design. After a pivot, it now stacks DRAM, the most common type of memory, directly on top of each chip, hoping to bypass the usual trade-off between memory capacity and access speed. That trade-off matters more in the agent era, as companies want large models to run long chains of tasks without sacrificing speed. Fractile also builds its own machine learning compiler and frames itself as a full-stack frontier AI company rather than just a chipmaker.

   CEO Walter Goodwin wrote on LinkedIn that the funding will let the company ship its first chips and systems next year, adding there is "still a lot of work ahead". Its chips are not yet shipping, so the performance advantages remain to be proven in products.

5. 05

   An agent model that fits on a phone: Tether open-sources 0.8B BotPsy, claims it beats models 2.5x its size at tool use

   Tether AI Research released QVAC BotPsy, an 800M-parameter open-weight model that turns requests into tool calls and can run locally on personal devices. In Tether's tests it roughly doubles its Qwen3.5-0.8B base on benchmarks such as BFCL v4 and τ-bench and outperforms Qwen3.5-2B.

   Read the full report

   Want an AI assistant to reschedule meetings and send messages without handing your email and calendar to the cloud? Tether's AI research team released QVAC BotPsy on November 11, an open-weight model with just 800 million parameters that turns a request into the right tool calls and is small enough to run locally on personal devices.

   BotPsy is fine-tuned from Alibaba's open-source Qwen3.5-0.8B-Think. In Tether's own testing it roughly doubled the base model on four main agent evaluations: BFCL v4 from 23.4 to 44.0, ACEBench from 30.1 to 58.7, τ-bench from 23.2 to 51.5 and ToolSandbox from 30.5 to 64.6, outperforming Qwen3.5-2B-Think, a model 2.5 times its size. On BFCL's multi-turn tool-calling test it scored 46.1 versus 23.8 for the 2B model. Tether also highlights a score of 86.1 on hallucination tests that check whether a model avoids calling a tool when none fits, the highest among models tested.

   In one example, a message failed to send because Wi-Fi was off and the inbox was full; BotPsy turned Wi-Fi on and freed space before sending, while the base model kept retrying. Tether's longer-term vision is a phone-based companion that hands complex tasks to a home computer or trusted peer-to-peer nodes.

   Weights and technical documentation are on Hugging Face. All results come from Tether's tests, and the model only produces tool calls; which services it connects to and how permissions work is up to developers' apps.

6. 06

   Rox, maker of AI sales agents, raises $100M at a $1.3B valuation a year after launch

   Rox, a startup building AI agents for sales, raised a $100M Series B co-led by General Catalyst and March Capital at a $1.3B valuation. Its agents automate customer research and outreach; Fortune framed it as betting it can take on Salesforce.

   Read the full report

   AI agents are moving into enterprise sales. Startup Rox raised a $100 million Series B on November 11 at a $1.3 billion valuation. According to Techmeme, citing Fortune, the round was co-led by General Catalyst and March Capital; Fortune's headline described Rox as an AI sales startup betting it can take on Salesforce.

   Rox team member @ishanmkh announced on X that the company reached the valuation just 12 months after launch, with Sequoia, Databricks Ventures, p7ventures and Factory also participating. Rox says it created the "Revenue Agent" category: AI agents that automate sales work such as customer research and outreach.

   The idea is to free salespeople from repetitive research, prospecting emails and lead follow-ups so they can spend more time with customers. The company did not disclose revenue or customer numbers, and claims about product impact come mainly from Rox itself.

7. 07

   Artist Mariko Mori used Claude to write code for a new video installation in her retrospective

   Anthropic's Claude account said Japanese artist Mariko Mori used Claude to write the code for a computer-generated animation, shown as a new video installation in her retrospective "All That Shines". The post drew thousands of likes.

   Read the full report

   AI coding tools are showing up in museums, not just software companies. Anthropic's official Claude account posted on November 12 that Japanese artist Mariko Mori, who has pushed the boundaries of art and technology for three decades, used Claude to write code for a computer-generated animation in her retrospective "All That Shines", presented as a new video installation.

   The post included a video clip of the work. The case shows how artists without deep graphics programming skills can use AI to turn visual ideas into runnable generative code while keeping control of the final result.

   The post did not explain how Mori worked with Claude or how much the code was edited by people, so those details rest on the official description.

8. 08

   Let AI pick the model: Rayline launches ARC router, claiming ~97% quality at a fraction of the cost

   Startup Rayline launched ARC, a routing model that picks the right model and reasoning level for each turn. Its founder claims 97.5% of Opus 5.5 quality at up to one-tenth the cost and says it is fully open source; the site lists integrations with Claude Code, Codex, Cursor and others.

   Read the full report

   Agents get expensive partly because even simple steps call the most expensive model. Startup Rayline launched ARC on November 11, billed as a unified model for routing AI: for each turn it decides which model should handle the task and with how much reasoning.

   Announcing it on X, the founder cited 97.5% of Opus 5.5 quality at up to one-tenth the cost and called ARC fully open source. Rayline's website uses slightly different figures: 97.4% quality retained at 14.3% of Opus 5.5's inference cost, or 7.1% when combined with local models. The site links a technical report on its methodology and benchmarks.

   ARC can be used like a regular model in tools such as Claude Code, Codex, Cursor, OpenCode, Cline, OpenRouter and llama.cpp. Rayline can run inference and return a response, or return just a recommended model and reasoning level for teams to run on their own gateway, cloud or local models. The site also offers a Decisions API for scoring and choice questions.

   Quality and cost figures come from Rayline's own tests, and actual savings depend on the workload; the scope of the open-source release should be checked against the official repository and license.

9. 09

   ElevenLabs launches an enterprise AI app marketplace with Salesforce and Zendesk; purchases via commitments next year

   Voice AI company ElevenLabs launched the ElevenLabs Marketplace at its New York summit, gathering industry voice agents built by partners on its models. Launch partners include Salesforce, IBM, NiCE, Twilio and Zendesk across finance, insurance, healthcare and retail. The CEO said enterprises will be able to buy them with ElevenLabs commitments from next year.

   Read the full report

   Voice AI company ElevenLabs is moving from selling models to building a platform. At its New York summit on November 11, it launched the ElevenLabs Marketplace, a curated set of industry AI and voice applications built by partners on its models, where enterprises can find agents that fit their workflows and connect with the companies behind them.

   Launch partners include Salesforce, IBM, NiCE, Twilio and Zendesk, covering financial services, insurance, healthcare, telecom, retail and government. Examples include Salient automating borrower servicing calls, Assort Health handling scheduling and inbound patient calls, Zendesk's agent resolving customer service requests end to end, and Consio's voice and SMS agents for e-commerce brands in 70 languages.

   CEO Mati Staniszewski said enterprises can find agents that already fit their industry workflows instead of building from scratch, and that from next year they will be able to buy them using their ElevenLabs commitments. The company also said enterprises now account for more than 55% of its revenue. For now the marketplace is mainly for discovery and introductions, with purchasing to come next year.

10. 10

    Catch parsing errors before agents do: Nace AI launches Drex for document processing at $7 per 1,000 pages

    Nace AI launched Drex for Document Processing, which uses its decision model to check and fix parsing errors in scans and photos of documents before they reach downstream agents. Nace says it ranks first on its own leaderboard with 38% faster corrections on RealDocBench; TestingCatalog reports pricing of $7 per 1,000 pages.

    Read the full report

    When agents process contracts or invoices, a common failure happens at step one: a scan or phone photo is parsed incorrectly, and everything downstream inherits the error. Nace AI launched Drex for Document Processing on November 11, designed to catch and fix those errors before parsed output reaches agents.

    Nace's post says Drex uses decision-powered parsing, ranks first on its own parsing leaderboard and beats GPT-6 Astra, Claude Opus 4.7 and every parser tested, with 38% faster corrections on RealDocBench. According to AI news account TestingCatalog, Drex can analyze and fix scanned pages or photos of documents, correcting a page 38% faster than regenerating it from scratch in Nace's tests, and costs $7 per 1,000 pages.

    The rankings and speed figures come from Nace's own leaderboard and tests, with no third-party evaluation yet. For teams handling lots of paper documents, it is available to try in Nace's console.

11. 11

    Jensen Huang resigns as Nvidia CEO; CFO Colette Kress to become interim CEO

    Nvidia co-founder Jensen Huang resigned as CEO on November 12 after reports of two affairs, including one with a woman suspected of spying for China, according to Bloomberg and other outlets. The Information reports finance chief Colette Kress will become interim CEO. Global AI stocks fell on the news.

    Read the full report

    What was a question of whether the board would oust him a day earlier became fact on November 12: according to Bloomberg, Reuters, the Financial Times and The Information, Nvidia co-founder Jensen Huang has resigned as CEO. Nvidia is the most important supplier of AI compute, and Reuters called his exit a big blow to the AI boom.

    Bloomberg reported that Huang resigned after reports of two affairs, including one with a woman suspected of spying for China. Nvidia's board had earlier opened an investigation into whether he was targeted by a foreign intelligence operation; Beijing previously called the espionage allegation groundless speculation. According to The Information, finance chief Colette Kress will become interim CEO. Kress joined Nvidia as EVP and CFO in 2013, and Bloomberg wrote that her job now is to stabilize the company.

    Markets reacted quickly. Reuters reported that global AI stocks sank as the exit deepened concerns about the sector, while CNBC looked at what it means for Wall Street's roughly $500 billion AI financing bet. Reuters separately reported that the crisis has AI firms courting former spies as security risks come into focus.

    Public information so far comes from media reports. A permanent CEO, the outcome of the board's investigation, and the longer-term impact on Nvidia's business and export-control policy remain to be seen. Nvidia is scheduled to report earnings on November 17.

12. 12

    Researcher says model weights carry a 'GPU fingerprint' showing DeepSeek, Kimi and others trained on restricted Blackwell chips

    Will DePue published an analysis arguing that initial model weights carry a fingerprint of the GPU used to create them. He says six open-weight models from DeepSeek, Moonshot, MiniMax and Ant Group were initialized on Nvidia Blackwell systems, which are barred from export to China, and that he reproduced results on real hardware. No company responses appear in the materials reviewed.

    Read the full report

    Can an open-weight model reveal which chips trained it? Researcher Will DePue said on X on November 12 that it can, and used the method to make a sensitive allegation: that the newest models from DeepSeek, Moonshot, MiniMax and Ant Group were built on Nvidia Blackwell chips, which are barred from export to China.

    His idea is that when a model is first set up, its starting weights are laid out in a pattern that depends on the exact GPU and seed, creating a hardware fingerprint whose traces survive training into the published weights. DePue says he searched more than six trillion combinations of chip layout and seed against open weights and found that six models, including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3, were initialized on Blackwell systems, four of them matching Grace Blackwell GB200 and GB300 rack-scale parts.

    He says he verified this on real hardware: running unmodified PyTorch with the recovered seeds, a B300 reproduced DeepSeek-V4-Flash's starting weights bit for bit and a rented GB200 reproduced Kimi K3's to a correlation of 0.999995, while an H100 reproduced none. He says the method also matched 24 known open models to the hardware their developers reported. He estimates Kimi K3 and DeepSeek-V4-Pro each used 2,000 to 3,000 Blackwell GPUs, adding up to more than half a billion dollars of export-controlled chips across these labs.

    These conclusions and estimates come from DePue's own report. He says the results were independently verified by senior researchers at OpenAI and Anthropic, without naming them. The method points to chip type, not where the chips were located or how they were accessed, such as through overseas rentals. No responses from the four companies appear in the materials reviewed.

13. 13

    Four nonprofits sue Anthropic, OpenAI and three others, asking a court to freeze more advanced AI

    Public Citizen and three other nonprofits sued Anthropic, Google, Meta, OpenAI and SpaceXSI in San Francisco Superior Court on November 12 under California's Unfair Competition Law, asking a judge to bar development and deployment of AI beyond current levels and to stop AI from directing AI research. It is a complaint; no court order has been issued.

    Read the full report

    A lawsuit is trying to use consumer-protection law to put the brakes on frontier AI. On November 12, Irreplaceable Education Fund, Public Citizen, Heat Initiative and the Alliance for Secure AI sued Anthropic, Google, Meta, OpenAI and SpaceX Super Intelligence (SpaceXSI) in San Francisco County Superior Court, seeking to freeze development of more advanced AI systems.

    The plaintiffs' core argument is that the companies' own executives have warned that advanced AI poses catastrophic risks, including enabling bioweapons and cyberattacks and escaping human control. The complaint says existing, less powerful systems have already bypassed safeguards and accessed outside systems without authorization. The groups invoke California's Unfair Competition Law, arguing these practices threaten public safety.

    There are two main requests. According to Edit USA's reading of the complaint, the first would bar each defendant from developing or deploying systems more capable than its most capable system at the time of filing; it does not expressly call for taking existing chatbots offline, but could affect future versions of ChatGPT, Claude and Gemini if granted as written. The second would bar AI systems from independently conducting or directing research on more capable AI.

    These are requests, not rulings; a judge will decide whether to grant relief and what it covers. The AP noted the suit follows a midterm election in which Democrats who won back Congress stressed the need to regulate AI. Responses from the defendants do not appear in the reports this article relies on.

14. 14

    Google launches Gemini 4 Hydrogen: 2.5x faster than 3.8 Flash, but pricier

    Google released Gemini 4 Hydrogen on November 12 for fast coding, multimodal and everyday agent tasks. Google says it finishes tasks 2.5x faster than Gemini 3.8 Flash and scores 51 on the Artificial Analysis Intelligence Index, edging Gemini 4 Argon. API pricing is $1.50 per million input and $7.50 per million output tokens; it is in the Gemini app, AI Studio and Cursor.

    Read the full report

    Google dropped the Flash name this time and launched Gemini 4 Hydrogen as its main model for everyday use and fast coding. Google says it follows instructions more closely, calls the right tools and reaches reliable output with less back-and-forth.

    Speed and capability are the pitch. Google says Hydrogen completes tasks 2.5 times as fast as Gemini 3.8 Flash at its default reasoning setting. On the Artificial Analysis Intelligence Index it scores 51 at default, ahead of Gemini 4 Argon's 50, and 54 at its highest xhigh thinking level; VentureBeat notes 3.8 Flash scored 31. It supports a 1M-token context and four configurable thinking levels. Cursor added it the same day, saying it scores 54.3% on CursorBench at default effort, up from 39.6% for 3.8 Flash.

    The trade-off is price. Standard API pricing is $1.50 per million input tokens and $7.50 per million output tokens, including thinking tokens, flat across the full 1M context; cached input is $0.075 per million, and Batch and Flex are half price. VentureBeat says that is about 15 times the lowest standard rates of budget models such as Claude Haiku 5.5 and GPT-6 Luna, and matches the new price 3.8 Flash moves to in January.

    For developers, Hydrogen may take on some work that would otherwise go to pricier reasoning models, but high-volume simple calls will need a fresh cost calculation.

15. 15

    Claude joins your video call: Claude Meetings builds docs and whiteboards while you talk

    Anthropic launched Claude Meetings in beta on November 12: teams can add Claude to a video call, and it builds docs, whiteboards and other artifacts with them while the conversation is still happening. It is starting to roll out on Team and Enterprise plans.

    Read the full report

    There are plenty of AI note-takers; Anthropic wants to go further by having Claude take part and produce work during the meeting itself. On November 12, the official Claude account introduced Claude Meetings, saying users can add Claude to their team's video call, where it builds docs, whiteboards and other artifacts (Claude's interactive outputs) with them while they are still talking.

    The difference from post-meeting summary tools is real-time output, such as drafting a document or sketching a whiteboard while a plan is being discussed. The announcement includes a roughly one-minute demo video.

    On availability, Claude Meetings is starting to roll out in beta on Team and Enterprise plans, so individual users cannot use it yet. The post itself does not list supported meeting platforms or permission and privacy details, which organizations should check before enabling it. The post drew more than 30,000 likes on X.

16. 16

    Thinking Machines opens Inkling 2 Small weights: a 594B-parameter MoE under Apache 2.0

    Thinking Machines released Inkling 2 Small's weights on Hugging Face under Apache 2.0 on November 12. The multimodal MoE model has 594B total and 17B active parameters, accepts text and images, and supports a 1M-token context; running it locally takes roughly 500GB of GPU memory.

    Read the full report

    AI lab Thinking Machines put Inkling 2 Small's weights on Hugging Face on November 12 under the permissive Apache 2.0 license, so developers can download it, fine-tune it and use it in commercial products.

    Despite the name, the model is not small. According to its model page, it is a 48-layer sparse mixture-of-experts (MoE) model with 594B total parameters; each token is routed to 8 of 384 experts, activating about 17B parameters, which keeps inference far cheaper than a dense model of the same total size. It natively accepts text and images, outputs text, supports up to a 1M-token context, and is aimed at agents, tool use, coding assistants and RAG.

    On requirements, the model was trained in low-precision NVFP4. The recommended NVFP4 checkpoint needs at least about 500GB of aggregate VRAM, such as 2x B300 or 4x H200; the MXFP8 version needs about 800GB. For most developers, APIs are more practical: the page lists providers including NVIDIA, Fireworks, Together, Nebius and Cloudflare, and inference frameworks such as SGLang and vLLM support it from day zero. SGLang says it reaches 576 tokens per second per user on NVIDIA Vera Rubin.

    The model card advises deployers to run their own performance and safety evaluations and to add validation and human oversight in high-stakes areas such as medicine and law.

17. 17

    Open-source inference engine SGLang hits 1.0, with Rust hot paths and caching beyond GPU memory

    SGLang, the open-source LLM inference framework, released v1.0 at the first SGLang Summit, bringing zero-overhead scheduling, Rust-rewritten critical paths, speculative decoding, hierarchical caching and prefill/decode disaggregation. It says it is used by 330+ organizations with day-zero support for 217 models. Miles v0.2 for RL also launched.

    Read the full report

    SGLang, an inference framework many companies use to serve open models, released version 1.0 at the first SGLang Summit in San Francisco on November 12, billing it as built for production at scale.

    The update has three parts. Faster inference: zero-overhead scheduling, performance-critical paths rewritten in Rust, speculative decoding (a small model guesses upcoming tokens that are then verified, to speed things up) and megakernels. Better memory use: unified memory for hybrid models, plus hierarchical caching that keeps long agent histories reusable beyond GPU memory, which matters for long-running agents. Flexible scaling: prefill/decode disaggregation and flexible parallelism so serving setups can match workloads.

    The project says SGLang is used by more than 330 organizations and had day-zero support for 217 models. LMSYS also introduced Miles v0.2 for production reinforcement learning. According to attendee Richard Chen, one slide showed that about 95% of merged PRs in SGLang are now written by agents, versus mostly by hand a year ago. Full release details are coming later, the project said.