AI Daily · November 12 Thursday
2026-11-11 08:00 — 2026-11-12 08:00 (UTC+8)
- 01
Alibaba teases Qwen4, opening a preview of flagship Qwen4-Max
Alibaba's Qwen team announced on X on November 11 that "Qwen4 is coming," with flagship Qwen4-Max now in preview on Alibaba's Token Plan and early access in QwenWork, Qoder, the Qwen app and QwenStudio. Qwen says it targets complex coding, knowledge work and agentic tasks, but has not published parameters or benchmarks.
Read the full report
Qwen's next generation is starting to appear. Alibaba's official Qwen account said Qwen4 is coming, and the first model people can try is a preview of the flagship Qwen4-Max: Token Plan subscribers can call it directly, and QwenWork, Qoder, the Qwen app and QwenStudio offer early access.
According to Qwen, the generation brings deeper knowledge, richer world understanding and stronger coding and agentic capabilities, aimed at complex coding, knowledge work, front-end builds and professional software workflows. The emphasis is less on chat and more on getting multi-step work done.
The announcement only sets out direction: it gives no parameter count, benchmark results, formal release date or pay-as-you-go pricing, and "preview" means capabilities and interfaces may still change. Qwen invited users to test it on real problems; real-world performance awaits hands-on feedback and the full release. The post drew heavy reposting and discussion on X.
- 02
Baidu releases ERNIE 5.5 Flash Preview: 280B total, 14B active, weights due by Nov 25
Baidu announced ERNIE 5.5 Flash Preview on X on November 11, saying it uses a new architecture built for efficient inference, with 280B total parameters and 14B active per token, a 1M-token context window and generation above 200 tokens per second. Model weights will be released by November 25.
Read the full report
Baidu has added a fast, cost-focused member to the ERNIE family. ERNIE 5.5 Flash Preview uses a new architecture designed for efficient inference: 280B total parameters, but only 14B active per token, which keeps inference costs closer to a mid-sized model.
Baidu's headline specs are a 1M-token context window and generation speed above 200 tokens per second. It says the model is stronger at coding, 3D modeling, knowledge work and content creation and posts competitive scores on Terminal-Bench, AutomationBench and other benchmarks, but the post lists no specific scores and the claims have not been independently tested.
For developers the key detail is openness: Baidu says weights will be released by November 25. For now it remains a preview, and local deployment and license terms can only be confirmed once the weights are out.
- 03
MiniMax opens H3-Image weights: native 16MP images with up to 16 references
MiniMax released H3-Image, an image generation and editing model, with weights on Hugging Face. It outputs roughly 16-megapixel images natively without an upscaler and takes up to 16 reference images; MiniMax says a 4K image takes about 10 seconds. The weights use MiniMax's own community license rather than a standard open-source license.
Read the full report
MiniMax has released a high-resolution image model. H3-Image unifies text-to-image and image editing in one checkpoint and generates images of about 16 megapixels directly, from 4096x4096 squares to 6272x2688 21:9 frames, with no upscaler. MiniMax says a 4K image takes about 10 seconds, and the model accepts up to 16 reference images, aimed at ad creative, e-commerce, and game and 3D assets.
According to the model card, the core is a 33.1B-parameter diffusion transformer distilled to 6 sampling steps without classifier-free guidance, which is the main source of its speed. The text encoder is MiniMax-H3's Qwen3-VL encoder.
Local deployment is demanding: the BF16 model is about 68.65 GB plus a roughly 66.7 GB text encoder. MiniMax recommends vLLM-Omni, SGLang, diffusers or ComfyUI, on hardware including NVIDIA H200, RTX PRO 6000 and AMD MI350X. It is also available via the MiniMax API, MiniMax Code and Hailuo AI, with 50% off 16MP+ generations during launch week.
The weights are under the MiniMax-H3-Image Community License, and some uses require separate written authorization, so check the terms before commercial use. Speed and quality claims currently come mainly from MiniMax's own showcases.
- 04
Anthropic launches IPO roadshow at $70-$78 a share, valuing it up to about $1.4 trillion
Anthropic said on November 10 it has launched the roadshow for its IPO of 825 million Class A shares at $70 to $78 each, applying to list on Nasdaq as "ANT." CNBC, citing people familiar, reports pricing on Nov. 18 and trading on Nov. 19, with a fully diluted valuation of up to about $1.405 trillion at the top of the range.
Read the full report
The AI industry's most watched listing is entering its final stretch. Anthropic's announcement says it has launched the roadshow for its IPO of Class A common stock: 825 million shares, about 801 million from the company and about 23.65 million from existing holders, at an expected $70 to $78 per share, with a Nasdaq Global Select Market listing under "ANT." Underwriters get a 30-day option on up to 123.75 million more shares.
According to CNBC, at the top of the range Anthropic would net up to $64.35 billion and could be worth up to $1.405 trillion fully diluted, up from the $965 billion valuation announced in its May funding round. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC have agreed to buy $26 billion of shares in a private placement at the IPO price.
CNBC, citing people who requested anonymity, reports pricing on Nov. 18 and trading on Nov. 19, with about 20% retail allocation versus the typical 5%-10%. The Verge, citing Reuters, notes the IPO had previously been expected to value the company at up to $2 trillion. Pricing and timing remain subject to final announcements, and the registration statement is not yet effective.
- 05
Google Cloud launches Gemini for healthcare, agents that work across records, coverage and billing
Google Cloud launched Gemini for healthcare on November 10 in U.S. preview. Built on the universal Gemini agent, it connects electronic health records, billing and payer policy systems to prepare charts, assemble prior authorization packets and check coverage. Launch customers include HCA Healthcare, Northwell Health and Optum.
Read the full report
Google wants AI to take over some of healthcare's most labor-intensive paperwork. In a blog post, Google Cloud CEO Thomas Kurian said Gemini for healthcare builds on the universal Gemini agent introduced last month and runs multi-step tasks across EHRs, billing platforms, clinical guidelines and CMS datasets without staff switching between apps.
For providers, it can compile a patient's history across care settings, prepare charts before visits and assemble authorization packets against payer policy before submission to reduce denials. For payers, it can check requests against medical policy and explain benefits and costs to members from their own plan documents. Google cites clinicians spending nearly 28 hours a week reconciling information by hand.
According to Fierce Healthcare and Google's press release, it connects with systems including Epic and Oracle Health, and launch or testing customers include HCA Healthcare, Northwell Health, Highmark Health, Optum and Stanford School of Medicine. It is for Gemini Enterprise Standard and Plus customers, with per-seat or pay-as-you-go pricing. Google says it supports HIPAA compliance, keeps data in the customer's environment, does not use it to train foundation models and keeps humans in the loop; real-world results remain to be shown during the preview.
- 06
Editing fewer than 10 neurons can backdoor an AI agent, Microsoft Research shows with Lazarus
Microsoft Research on November 10 disclosed Lazarus, a backdoor technique that, without retraining, rewrites fewer than 10 rarely active neurons so an agent behaves normally and passes standard safety evals until a trigger makes it run malicious commands. Microsoft says model provenance and weight integrity are critical.
Read the full report
An open model that looks normal and passes benchmarks could start working for an attacker after a certain date. Microsoft principal researcher Dimitris Papailiopoulos writes that reprogramming the weights of fewer than 10 dormant neurons in an LLM is enough to backdoor an agent built on it, with no training; any machine with enough RAM to run the model can do it.
The technique turns rarely firing neurons into trigger detectors. The trigger can be mundane, such as the clock passing midnight into 2027, at which point the agent, mid-task, emits a command that downloads and runs attacker code with whatever access it has. For agents with terminal or tool access, that could mean quietly stealing secrets or running destructive commands.
The good news, the researchers say, is that such edits can be caught: frontier agents can spot anomalous weight modifications when asked, but reliable detection requires diffing against the original weights, so provenance matters. They stress this is not a case against open weights, since internal models are equally exposed through compromised accounts, pipelines or storage. Microsoft says Foundry protects customers via sourcing controls and is developing weight scanning, red-teaming tools and tamper-evident model signing. The author says Claude Code discovered most of the findings autonomously.
- 07
Google DeepMind releases Gemini Robotics 2.5; humanoid hits 96% success sorting parts on Mercedes lines
Google DeepMind released Gemini Robotics 2.5 on November 10, a vision-language-action model that it says achieves 30%-100% higher success and 20%-30% faster completion than prior baselines across 200+ tasks. In a proof of concept with Apptronik and Mercedes-Benz, Apollo 2 humanoids sorted automotive parts on production lines with 96% success.
Read the full report
DeepMind has moved robots another step from the lab toward the factory. Gemini Robotics 2.5 is its most capable vision-language-action (VLA) model, one that sees, understands instructions and directly controls robot arms. Across 200+ tasks such as serving food with tongs, tying a bow on a gift box and pipetting liquid, DeepMind says success rates rose 30% to 100% over prior baselines and tasks finished 20% to 30% faster.
The headline is a real production test. With Apptronik and Mercedes-Benz, Apollo 2 humanoids performed automotive kitting on lines in Kecskemet, Hungary, and Sindelfingen, Germany, sorting metal brackets from bulk bins into sequenced trays with a 96% success rate. It remains a proof of concept rather than a large-scale deployment.
On safety, the model steers around obstacles and nearby people at inference time, which DeepMind says raised safe task completion by 33% and 43% respectively. It also open-sourced ASIMOV-Jailbreak, a robotics jailbreak benchmark, on which it says Gemini 4 Argon achieves 99.5% resistance. Argon also generated simulated practice runs from single camera views to speed training. The model is rolling out to robotics partners, not to general developers.
- 08
Grok Build share-link flaw exposes 35,000 enumerable apps, including SpaceXSI internal files
Ars Technica reports that security firm Claroc found Grok Build share links put their random words in the domain, letting anyone enumerate 43,601 links via passive DNS. More than 35,000 opened directly, exposing internal business plans, staff rosters and medical exams, including data from 24 apps built by SpaceXSI itself.
Read the full report
An internal tool built with AI in minutes can be visible to the whole internet the moment it goes live. According to Ars Technica, security firm Claroc found a design flaw in app share links for Grok Build, the vibe-coding agent from SpaceXSI (the merged SpaceX and xAI), that exposes large amounts of user data.
The issue is the link structure. Grok Build generates addresses like "xxx-xxx-xxx.grok.me" from random words so only link holders can access them, but because the randomness sits left of the domain, first visits leave DNS lookups that are publicly searchable in passive DNS databases. Claroc says it found 43,601 grok.me links, more than 35,000 (nearly 95%) of which opened the app and all its data. Google Docs and the share links of Claude, ChatGPT and Gemini keep randomness in the path, invisible to DNS.
Exposed data included medical exams for 3,784 workers at a Peruvian construction firm, a Tesla service-center technician scorecard and a roster for a U.S. Navy contractor. SpaceXSI's own 24 apps were among them, including a confidential sales onboarding plan and a 594-account territory map; those links were taken down within hours of Ars's inquiry.
A Cursor representative said Grok Build apps are private by default and nothing goes public unless the creator shares it, but Claroc's tests found the default was "Anyone with the link." Ars notes Lovable, Bolt and Replit app addresses carry similar DNS exposure. Users should switch sharing to invited users only.
- 09
Tencent launches Yuanbao Search, an AI search engine covering WeChat articles and Channels with collapsible ads
Tencent launched Yuanbao Search for consumers on November 11, available at so.qq.com or in the latest QQ Browser. Built on the Yuanbao large model with an index of hundreds of billions of pages, it covers the web, WeChat Official Accounts, WeChat Channels and academic papers, and separates ads from answers with an option to collapse them.
Read the full report
Tencent is entering web search directly. On November 11 it launched Yuanbao Search, available by typing so.qq.com in any browser or by switching the search engine in the latest QQ Browser.
Tencent says Yuanbao Search uses the Yuanbao model to rebuild the whole pipeline from understanding queries to retrieval, ranking and composing answers. Its index of hundreds of billions of pages covers not only the web but also WeChat Official Accounts, WeChat Channels and academic literature. That matters for Chinese users, since WeChat articles have long been hard to find in general search engines.
Results prioritize authoritative and official sources, drawing on 157 such sources, with an AI summary at the top and a one-click "AI mode" for follow-up questions. Papers get plain-language explanations with original citations and one-click translation of English texts, and users can search by typing, voice or photo. Tencent stresses that ads will not masquerade as answers: they sit in a separate, labeled block that can be collapsed.
According to 100EC, product lead Li Ruizhang said Yuanbao Search will not evolve into an AI chat assistant, is positioned differently from Sogou, and is entering web search as a challenger.
- 10
Inkbox raises $5.3M and launches hosted agents with their own computer, email and phone
Ray Liao of agent communications platform Inkbox announced on X a $5.3 million raise led by Peak XV and the launch of hosted agents: each runs on its own computer alongside email, phone, iMessage and WhatsApp channels, and sleeps when idle to save costs.
Read the full report
If AI agents are to send messages and make calls like coworkers, they need an always-reachable but affordable place to run. That is Inkbox's pitch. Inkbox's Ray Liao announced on X that Inkbox raised $5.3 million led by Peak XV, with several tech industry angels participating.
Alongside it, Inkbox launched hosted agents: an agent runs in Inkbox's compute environment next to its email, phone, iMessage and WhatsApp channels, so it never misses a reply, and it sleeps when idle so users don't pay for downtime. According to the post, the platform has processed millions of messages for AI agents in the four months since launch.
The post passed 1.8 million views on X and prompted discussion about agent permissions: once AI can message and call people on its own, whom it represents, what it may do and whether sends need confirmation matter as much as model capability. Features and figures are the company's own claims.
- 11
Google launches a choice-only AI: Gemini Decision 1 answers up to 1,000 questions in one pass, with free output
Google launched the Gemini Decisions API in AI Studio, powered by the new Gemini Decision 1 model. Instead of generating text, it answers up to 1,000 yes/no, multiple-choice or rubric questions about one multimodal input in a single pass. Google cites 150-250 ms latency and $0.10 per million input tokens, with output not billed.
Read the full report
Google has released a model that doesn't write prose, only makes decisions. The new Gemini Decisions API, powered by Gemini Decision 1, handles bounded choices: whether content breaks a policy, which tool to call, which model a request should be routed to, or how to grade a piece of context. Rather than generating an answer token by token, it scores candidate options in a single forward pass and returns probabilities plus a confidence score from 0 to 1.
According to Google's blog, one request can carry a shared input of up to 256K tokens mixing text, images, audio, video and PDFs, plus up to 1,000 independent questions. There are three question types: yes/no, multiple choice with 2 to 255 options, and ordered rubric scoring. Google cites 150-250 ms end-to-end latency and pricing of $0.10 per million input tokens, with zero output tokens billed and a no-cost tier in AI Studio for prototyping.
The target is the inside of agents, where complex tasks involve many small branching decisions; calling a general chat model and parsing its text each time is slow and expensive. Confidence scores also let developers route uncertain cases to fallbacks or human reviewers.
For now it is available only on the Gemini Developer API v1beta and requires google-genai 3.2 or later. Unite.AI reports that question instructions are text-only and that conversation history, context caching and the Batch API are not supported; Vertex support on Google Cloud is listed as coming soon. Speed and cost claims are Google's own, pending developers' real-world testing.
- 12
Same DNA, different cells: DeepMind's AlphaGenome 2 predicts variant effects by cell type
Google DeepMind released AlphaGenome 2, trained on more than 100 million single-cell profiles, to predict how DNA variants act in specific cell types. DeepMind says it finds nearly twice as many causal variants as its predecessor on an immune-cell dataset. It is available via API for non-commercial research.
Read the full report
Nearly every cell in the body carries the same DNA, but heart, brain and immune cells read it differently, so a single-letter change may cause disease only in certain cells. DeepMind's new AlphaGenome 2 aims to compute that difference: given a DNA sequence and a target cell type or state, it predicts how variants change gene activity.
Unlike the first model, which predicted from bulk tissue averages, AlphaGenome 2 is a conditional model that predicts for one cell type or state at a time. DeepMind says it was trained on bulk and single-cell data from humans and mice, including more than 107 million individual cell profiles across 68 tissues. On the TenK10K dataset of 28 immune cell types, it identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data, and direction-of-effect recall rose from 23% to 43%.
DeepMind also says AlphaGenome 2 computationally reproduced a recent finding that a distant regulatory variant drives inflammatory bowel disease by overactivating ETS2 in gut macrophages. Researchers can upload their own single-cell RNA-seq data to adapt the model to unseen cell states without retraining.
Nature quoted scientists saying the first model's inability to resolve cell types limited its usefulness, the gap the new version targets. AlphaGenome 2 is available through the AlphaGenome API for non-commercial research and in private preview for commercial research. Performance figures come from DeepMind's blog and preprint and await further independent validation.
- 13
Video company Runway open-sources a robot model: Praxis-1 weights released under Apache 2.0
Runway made its robotics world action model Praxis-1 available to everyone with open weights under Apache 2.0, including inference and fine-tuning code, for commercial use. The 27B model scores 46.2% on NVIDIA's RoboLab-120, ranking first among world action models and open-source robot foundation models.
Read the full report
Runway, best known for AI video generation, is bringing its video models to robots. The company made Praxis-1 available to everyone, with weights, inference code and fine-tuning code under the Apache 2.0 license, including for commercial use. It is a world action model: before each move, it predicts both what the scene will look like next and the motion needed to get there.
According to Runway's research page, Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer adapted from a large video pretrained model, with a small action expert that reads the video model's internal state. On NVIDIA's RoboLab-120 simulation benchmark (120 tabletop tasks, 1,200 episodes), the DROID fine-tune succeeded in 555 episodes, or 46.2%, ahead of FLUX 3 Action at 42.9% and NVIDIA Cosmos3 Nano at 36.8%. Note that it ranks fifth overall on the leaderboard; "first" refers to the world action model and open-source robot foundation model categories.
Runway released three checkpoints, a single-arm DROID version, a bimanual YAM version and the base model, with support for full and LoRA fine-tuning. It says the model runs on hardware from H100 and B200 to RTX 5090 and on-robot Jetson Thor, and an NVFP4 version shrinks policy weights from 54.4 GB to 17.5 GB. Launch partners including Noble Machines and Standard Bots have run it on a humanoid, an industrial arm and a warehouse robot.
The GitHub instructions show it also needs components such as Wan2.2 and Qwen3.5, which carry their own licenses, so commercial users should check each one.
- 14
A European military intelligence AI: Helsing launches Clarity, using hundreds of agents on satellite and drone data
German defense tech firm Helsing launched Clarity, an agentic intelligence platform for armed forces. Helsing says hundreds of AI agents analyze satellite imagery, drone video, signals and field reports in parallel, marking adversary assets with evidence for human review. It stresses that every engagement is a human decision and that forces own the infrastructure.
Read the full report
AI agents are moving into military intelligence. German defense tech firm Helsing launched Clarity, an agentic intelligence platform "built in Europe, under European control." The problem it targets: satellites capture huge volumes of imagery daily and one reconnaissance flight can record more than a billion radar pulses, yet turning data into action remains largely manual.
According to Helsing's site, hundreds of agents watch satellite imagery, drone video, signals, open sources and field reports, flagging changes to a master agent that builds a shared operational picture. Findings gain weight only when independent sources agree, and every conclusion traces back to evidence. The system also recommends a suitable effector, such as a missile or autonomous platform, and checks whether the data is precise enough.
Helsing stresses human control: a human nominates every threat and authorizes every hand-off, weak or outdated evidence is stopped, and legal review and rules of engagement are built in with every decision recorded. Clarity follows NATO standards, connects to effectors from any manufacturer and runs on infrastructure the force owns.
Several market accounts framed Clarity as a European rival to Palantir's Maven; that is outside commentary, not Helsing's wording. The launch also revives the debate over whether "humans decide" holds up once AI takes part in targeting, which will depend on real deployments and oversight.
- 15
OpenAI research: smarter models can reason without writing it down, making chain-of-thought monitoring harder
OpenAI researchers published work on how chain-of-thought monitorability evolves over training. Models with stronger pretraining solve more tasks without visible reasoning and better control their CoT when prompted; reinforcement learning substantially raises that controllability, and it can be distilled into the next model generation.
Read the full report
Having AI write out its reasoning has been seen as a key window for spotting misbehavior. A set of results OpenAI researcher Jeffrey Wang shared on X suggests that window may narrow as models get stronger, findings the team called "unsurprising but sobering."
The work tracks two measures. No-CoT capability is the ability to solve tasks in a single forward pass without written steps; the stronger it is, the harder it is to see how a model reached an answer. CoT controllability is the ability to add, remove or manipulate chain-of-thought tokens when prompted. Across internal and released frontier models, no-CoT capability improved smoothly as base model perplexity fell (stronger pretraining), and controllability showed a weaker correlation. On multi-hop planning and serial computation, some architectures beat what perplexity predicts.
Training matters too. In an "Astra-class" model, CoT controllability rose substantially during RL even with only length constraints and final-answer grading; adding tool access and tool supervision raised it slightly. The team also found controllability can be passed to the next generation through supervised fine-tuning on model-generated data.
The researchers note that for two similarly capable models they saw no meaningful link between controllability and monitor evasion, calling controllability a noisy proxy for adversarial evasion. Details are in OpenAI's blog; the conclusions are the team's own.
- 16
Is business AI spending cooling? Ramp: October AI spend rose just 2%, with OpenAI and Anthropic growth in low single digits
Ramp's October AI Index shows business AI spend up just 2%, with OpenAI and Anthropic growth falling to low single digits and engineering teams driving 72% of spend, according to its lead economist. Business AI adoption rose to 56.8%, and Anthropic still leads token spending share.
Read the full report
Business billing data is showing signs of cooling in AI's growth story. Ramp lead economist Ara Kharazian posted on X that the latest Ramp AI Index shows business AI spend rose just 2% in October, with OpenAI and Anthropic growth down to low single digits; engineering teams drove 72% of spend, followed by customer service and leadership.
Ramp's data covers more than 85,000 U.S. businesses using its corporate cards and bill pay. The index page shows paid AI adoption at 56.8% in October, up 1 point from September and 13.3 points from a year earlier, with Anthropic at 44.8%, OpenAI at 40.6% and xAI up from 4.1% to 6.3%. The gap between heavy and typical users is stark: among paying firms, the top 1% had a median of about $10,199 per employee per month, versus $16.32 for the median firm.
A separate dataset from Ramp's token spend management product shows Anthropic at 49.3% of token spend in the week ending November 1, down 3.9 points from a month earlier, with OpenAI up to 46.1%. Their blended price fell about 15% in a month to roughly $0.53 per million tokens, which may partly explain slower spend growth.
Mind the scope: Ramp sees only payments made on its platform, not free tools, personal accounts or large direct enterprise contracts, and it acknowledges likely underestimation. It is better read as a trend signal than as a measure of AI companies' total revenue.
- 17
SemiAnalysis founder: 45% of the firm's AI spend now goes to persistent agents, and Claude and OpenAI agent products fall short
SemiAnalysis founder Dylan Patel said on X that 45% of the firm's AI spend now goes to persistent agents, mostly Perplexity Computer, and criticized Anthropic's Claude Tags and OpenAI's Slack agent. He added that Perplexity mainly runs Anthropic models for the firm's use cases.
Read the full report
Persistent agents, AI that stays online and keeps working on tasks rather than answering one prompt at a time, are taking a real share of corporate AI budgets. Dylan Patel, founder of semiconductor research firm SemiAnalysis, said on X that 45% of the company's AI spend now goes to persistent agents, calling them the next unlock of AI usage.
His product verdicts drew more attention. Patel said the firm primarily uses Perplexity Computer, which draws on many models, and praised Perplexity's harness; he said Anthropic's Claude Tags "sucks" and OpenAI's Slack agent is "even worse." In a follow-up he added that Perplexity mainly uses Anthropic models for their use cases, but argued that Anthropic and OpenAI are bad at product and not receptive to feedback.
The remarks sparked debate over whether model leaders will build the best agent products or whether orchestration layers on top will win users. These are one executive's views based on his firm's own use, and the 45% figure came without a stated methodology, so it does not represent the industry as a whole.
- 18
Do agents really learn on the job? New Continual-ARC benchmark says OpenClaw and others don't
Unify's Daniel Lenton released Continual-ARC, a continual-learning benchmark, saying OpenClaw, Hermes Agent and Prime Agent failed to memorize anything of value or generalize to new tasks, while training a small 9B model's weights did both. The post passed 670,000 views and drew heavy debate.
Read the full report
Many AI agents advertise memory and learning from experience, but a new test throws cold water on that. Unify's Daniel Lenton announced the team's new benchmark, Continual-ARC, saying OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience: none memorized anything of value and none generalized to new tasks. Training the weights of a small 9B model, by contrast, did both.
As described in the discussion, the benchmark streams repeating and new tasks to test both remembering past work and generalizing. The key contrast is that most agents' "memory" means writing notes or summaries into context or external files while model weights stay unchanged; Lenton's results suggest that may not produce real capability gains.
The post drew more than 670,000 views and hundreds of quotes and replies. Many users complained about re-explaining the same things to agents daily, while some developers questioned whether the setup was fair and how long-term improvement should be measured. The benchmark was designed and run by Unify, which builds related products, so the results await third-party replication.
- 19
a16z: the internet's main users are no longer human, and the B2A economy is here
a16z published a thesis on the "B2A" (business-to-agent) economy, arguing AI agents now make up most web traffic, write over half of new online articles and generate most production code, and will increasingly decide what users buy. It cites Shopify, Stripe and Vercel data, but the core claims are an investor's view.
Read the full report
Businesses once built walls to keep bots out; soon they may compete for bots' business. a16z, in a long X article titled "Welcome to the B2A Economy" posted by Olivia Moore, argues the internet's primary users are no longer humans but AI agents: they make up most web traffic, write more than half of new articles published online and generate most code pushed to production.
As agents move from answering questions to completing tasks, the piece says, they will increasingly decide which products to use, which services to subscribe to and where to spend money. a16z cites third-party figures: Shopify said AI-driven traffic to its stores grew eightfold year over year in Q1 2026 and AI-search orders nearly thirteenfold; Human Security saw agentic browser traffic grow 8x from August to October; Stripe said agents account for nearly 40% of documentation traffic; Vercel said agent-triggered deployments rose 28x in a year; and by September, upward of 40% of GitHub PRs started from an agent's draft.
a16z treats coding agents as the template. Developers are under 1% of the global workforce, and when the other 99% use agents, payments, fraud, identity and communications infrastructure will be strained. It also concedes most enterprises still block agent traffic over fraud and chargeback concerns.
The claim that agents are the majority of traffic is a16z's judgment, the cited figures use different company-reported measures, and a16z has invested in many related companies.
- 20
Google Antigravity's agent can now drive a browser: research, code and verify live in one loop
Google's AI coding tool Antigravity says browser use is rolling out, letting its agent control the built-in browser for research, localhost testing and repetitive workflows, and run multiple tasks at once. Users choose which sites it logs into, and downloads pause until approved.
Read the full report
AI coding tools are taking over the step of opening a page to check the result. The official account of Google's AI coding environment Antigravity said browser use has been rolling out: the agent can control the built-in browser, closing the loop between researching, writing code and verifying it live.
The post lists a few details: users choose what the agent logs into, downloads pause until the user approves them, and the agent can use the browser to run multiple tasks at once. Per Google's demo, the feature suits localhost testing, online research and repetitive web workflows.
For developers, it means no longer manually refreshing, clicking and checking after a front-end change; the agent can run through it and spot problems. The risk is that browsers hold sessions and sensitive data, so the login and download approval design matters. The feature is rolling out gradually, so not everyone has it yet, and the post did not say whether limits differ by subscription tier.
GitHub projects
From GitHub Trending, with original project descriptions. Trending does not imply a new release.
-
addyosmani/agent-skills ↗
Production-grade engineering skills for AI coding agents.