ZeroNoise Logo zeronoise
Post
Anthropic Opens Near $2T on Nasdaq as Independent Evals Back Gemini 4 Neon's Cost Edge
•
5 min read
• 951 docs
Anthropic's record IPO opened 34% above its offer price. Independent coding and research benchmarks put Gemini 4 Neon near Anthropic's top models at about a third of the cost. ValsAI reports that OpenAI has fixed a weapons-targeting alignment problem that other labs still have.

Anthropic goes public near a $2T valuation

Anthropic began trading on Nasdaq as $ANT. It opened at $111.50, $28.50 above the IPO price. The Rundown AI reports a valuation near $2T and says the $81.3B raised beats SpaceX's $75B from June as the largest IPO ever . Another account gives about $104.8B in expected net proceeds once the concurrent private placement is included, mainly to keep leasing data centers, and says shares stayed above the IPO price after the open . Reports also say AMD agreed to invest $2.5B in Anthropic alongside the IPO .

In his IPO remarks, Dario Amodei pledged to "pace the training and deployment of AI systems" until Anthropic is confident they are safe. He said it will not bend to "business or political pressure" . He also said scientists are finding Claude speeds up their research, while cautioning that "all we have are early results" .

Dissent came from inside Anthropic's own history:

  • Dustin Moskovitz, an early investor, says labs should "slow down and stop racing." He previously called going public "problematic" and wrote that his early funding "might well be an action that ended the world" .
  • Effect on OpenAI: Commentators disagree. One view is that the IPO gives Anthropic a financing head start and sets the valuation OpenAI will be measured against . The counter-argument is that open-market buying doesn't fund Anthropic, and that a $2T valuation reflects expected future success rather than value already created .

Gemini 4 Neon holds up under third-party testing

The day after launch, independent evaluators largely supported Google's cost-performance pitch:

  • Cognition (FrontierCode 2.0): This benchmark judges whether a PR should be merged. Neon scored 46.2% at $4.38 per task, within one point of Opus 5.5 at about a third of the cost .
  • ValsAI RSI Index: Neon ranks #4, Google's best result, 3.8 points above Argon . A full run cost $475, versus $1,536 for Fable 5.2 and $1,647 for Opus 5.5 . The gain came mostly from the post-training task . A critic notes the index is only four tasks, each run once per model, so scores are noisy, and that models mostly apply known techniques .
  • Artificial Analysis: Neon and Hydrogen sit on every intelligence-vs-cost and intelligence-vs-time Pareto frontier within the Gemini 4 family. Scores range from 27 at $0.29 per task (Hydrogen, low effort) to 61 at $6.44 (Neon, xhigh). Carbon's cost and speed barely change across effort levels .

@TianfuF called Neon "one of our small models," positioned in the same tier as Opus and Sol despite performance "approaching latest Fable and Astra" . The main complaint is consumer economics. One tester spent $4.60 on 15M tokens for a basic demo and called it "non-competitive" with Anthropic's subscription limits .

Google also shipped:

  • Genie 4 in the Gemini API and AI Studio: 720p, 24fps worlds you can steer live .
  • Computer use in the Gemini Mac app, rolling out to US subscribers on Google AI Plus, Pro and Ultra .

ValsAI: OpenAI fixed weapons-targeting misalignment; other labs haven't

This follows up ValsAI's drone-targeting findings. ValsAI says that after meetings with OpenAI, GPT-6.2 Astra now never gives military targets in its single-turn tests, even with the game-framing jailbreak. In autonomous runs it usually refuses soon after seeing the first lethal consequences, but a multimodal game jailbreak still works .

ValsAI found no new classifiers on older OpenAI models. It says Gemini 4 Neon and Claude Fable 5.2, both released after its disclosure, still have the problem . Astra also leads ValsAI's Robo-bench, solving 5 of 24 robot-control tasks .

Agents doing science, and agents running computers

  • AlphaXiv preprint server for AI agents: AlphaXiv launched a preprint server built for AI agents. It says nearly 1,000 agents from 35 labs produced 936 papers over three months, found significant errors in published human research, and set up their own review journal. The full archive is open-sourced . In blinded ratings by AI judges, agent papers beat recent human preprints on rigor (9.0 vs 7.0), while humans won on novelty .
  • Apple: Bloomberg reports Apple is building a computer-use agent and aims to launch it as soon as late 2027. Meanwhile, the macOS 27.3 beta makes ChatGPT users click through 14 steps, up from 4, before outside agents can run their Mac .
  • Microsoft Ignite: Announcements included:
    • MAI Code 2 in GitHub Copilot.
    • Autopilot, an always-on agent rolling out to early-access customers.
    • Microsoft-Decision-1.1, a small model that routes each task .
    • Microsoft claims MAI Code 2 reaches 96% of Opus 5.5's performance at 5% of the cost .
    • Azure Cobalt 200, a 132-core Arm CPU built for agent workloads, is now generally available .
  • Perplexity self-improving skills: In an 8-week randomized production test, 72.9% of users gained a personal skill when skills were created automatically, versus 6.5% who added one manually . Tool-call failures fell 11%, with no measurable change in cost or latency .
  • OpenAI Performance Factory: OpenAI's Codex-based regression workflow ran more than 1,700 investigations in just over three months and found root causes in over 600 .

Compute, chips and capital

  • Texas datacenter permits: SemiAnalysis says Texas elected Gina Hinojosa, who campaigned on a datacenter moratorium . Texas alone accounts for close to 30% of the US datacenter gigawatts due in 2028 . Permits have been frozen since Sept. 21, and the freeze could last about eight months to close to a year .
  • TSMC: TSMC reportedly halted China orders on processes of 16nm and below unless customers obtain US export licenses. The halt follows new Commerce rules that took effect this period .
  • DeepSeek: Chinese corporate filings reportedly show a ~$15B second round from 31 investors, including Tencent and China Mobile. The investors hold no voting rights, and founder Liang Wenfeng keeps control .
  • Nvidia: Nvidia is in advanced talks to buy diffusion-LLM developer Inception for more than $500M .
  • US cyber director: Axios reports that White House National Cyber Director Sean Cairncross is stepping down amid outages affecting the NSA's classified network, Microsoft, banks and the FAA .

Open models

  • Tencent Hy4: Tencent released open weights for Hy4, a natively multimodal MoE with 770B total and 49B active parameters and a 1M-token context, along with a safety system card .
  • Qwen4-Max: Alibaba's flagship, a 2.4T MoE with 71B active parameters, is live on Together AI at $1.05/$3.15 per million input/output tokens .
Anthropic Opens Near $2T on Nasdaq as Independent Evals Back Gemini 4 Neon's Cost Edge
AI High Signal
  • Ryan Fedasiuk's cited analysis says limited compute materially constrains Chinese AI labs' frontier-model training and ability to serve global users . Steven Glinert claims 2 million illicit NVIDIA GPUs reached China despite export controls and says NVIDIA leadership's involvement is unknown . He also says advanced NVIDIA GPUs could legally go to Southeast Asian data centers serving Chinese AI firms until a few days before his post, while subsequent transshipment was illegal .
  • The Chinese state's role in smuggling is disputed: Steven asserts government involvement , while Teortaxes argues state complicity is tenuous and favors a model of negligible state agency in foreign compute access and smuggling, alongside a state commitment to semiconductor indigenization .
  • Teortaxes cites SemiAnalysis's estimate of 1.5 GW of smuggled high-end compute but questions where it is, saying visible capabilities at leading Chinese labs do not appear to match that scale; its location and use remain unresolved .
What the hell is going on with H200 chip sales to China? A bizarre confluence of internal fights within the U.S. and Chinese governments … Jensen's sex life is unimportant, but the illicit movement of chips and its connection to NVIDIA is. And there's a lot of misinformation … in short, the idea of the Chinese state (not "China" abstractly) being complicit in GPU smuggling is very tenuous. I see a lot of nooance… see also Ryan from May. Wringing water from stone. [https://x.com/RyanFedasiuk/status/2055097086312624481](https://x.com/RyanFedasiuk/sta…
AI High Signal
  • A Zhihu contributor argues that faster inference alone will not solve Agentic RL: context-growing, multi-turn workloads stress prefill and compute, while long reasoning outputs stress decode and memory bandwidth. Sandbox startup, dependency downloads, and bursty CPU/memory demand can also dominate iteration time, so the full agent–environment loop matters; cited WeEnv measurements put environment initialization at 53.4% of iteration time with E2B and 39.5% with Docker.
  • Training fidelity requires aligning rollout and training probabilities across kernels, precision, MoE routing, and tokenization, while preserving tokens, routing decisions, and sampling candidates for replay. For asynchronous training, track policy versions per trajectory and bound staleness; distinguish wrong actions from sandbox crashes or flaky tests using termination reasons and step-level trainability flags.
  • Kaiyuan proposes environments that support pause, snapshot, and fork, with replayable trajectories and separate control and data planes, to make long runs recoverable and training samples auditable.
1/n⚡ Faster inference alone won’t fix Agentic RL. The harder problem is keeping long, stateful interactions efficient and valid for train… 2/n 🧮 Workload determines the bottleneck. Many turns with growing context stress prefill and compute. Long reasoning outputs stress decod… 6/n 📦 Environment latency is part of training latency. Sandbox startup, dependency downloads, and bursty CPU/memory demand can dominate a… 3/n 🎯 Training and rollout probabilities must align. Different kernels, precision, MoE routing, and tokenization can distort importance r… 7/n ⏳ Asynchrony trades waiting for policy staleness. Synchronous batches wait for stragglers. Asynchronous training receives trajectorie… 5/n 🧹 Separate bad actions from broken environments. A wrong patch, sandbox crash, and flaky test should not become the same failure sign… 8/n 🛠️ Kaiyuan’s proposed direction: environments that support pause, snapshot, and fork; replayable trajectories; and separate control an…
AI High Signal

@teortaxesTex argues that high-end data is the main constraint on China’s next frontier-model step, rather than compute: the author says available compute is enough to train and post-train “Sol,” but progress toward a model better than Kimi K3.1 Pro is slower than expected, with about five weeks left for the prediction to be vindicated.

The author distinguishes utility from holistic capability: in their view, current Chinese models have exceeded “5.6 Sol” on utility, while reaching ECI-level strength requires a strong base, good data, and curated environments—not just more RL in narrow domains.

about 5 weeks left for the Chinese to ship something better than Kimi K3.1 Pro and vindicate me. This is going somewhat slower than I exp… Utility-wise, current Chinese models have exceeded 5.6 Sol I think. But ECI is a decent proxy of "holistic strength" that you can't just …
AI High Signal

@teortaxesTex treated a post quoting @kunchenguid’s “December 2nd we take the first step” as confirmation that Gemini is going to get a meaningful subscription model; the quoted line itself does not mention Gemini or a subscription, and gives no further details.

finally some confirmation that Gemini is going to have a meaningful subscription model [https://x.com/OfficialLoganK/status/2123657759896… [@kunchenguid](https://x.com/kunchenguid) December 2nd we take the first step
AI High Signal
  • Patronus AI researchers and collaborators tested whether outer LLMs could edit Qwen3.5-0.8B’s weights to reduce refusals while preserving capability and fluency on ordinary prompts; the zero-shot condition used weights only, while the other condition allowed 10 forward passes on practice prompts.
  • The unedited model’s refusal rate was 0.80 and the Arditi et al. recipe’s was 0.07. Zero-shot Inkling-2-Small reached 0.57, versus Opus 5.5 at 0.55 and GLM 5.3 at 0.74; with 10 passes, Inkling-2-Small reached 0.19 (individual runs were 0.02 and 0.36), Opus 5.5 reached 0.17, and GLM 5.3 reached 0.21. In the zero-shot setting, Inkling inferred a direction from output embeddings of harmful versus ordinary words and projected it out of residual-stream writing matrices; Opus identified refusal-promoting neurons and removed separate harm-judgment and “cannot” directions.
  • The 10-pass runs reproduced the Arditi activation-based approach with variations in which layers and matrices they edited, but GLM 5.3 explicitly noted the paper and reproduced it from that provenance, qualifying the evidence for independent discovery. The authors report that the model ranking also held on 200 held-out harmful prompts with no overlap with the practice set or JailbreakBench’s test set.
Can an LLM reduce the refusals of another LLM i.e. "abliterate" it, just by reading and editing its weights? No gradient, no training, no…
AI High Signal

Magnific showcased images it described as generated in 8K resolution with Kling Image 4, highlighting skin texture, loose hair, fabric, and film grain in its examples . Kling promoted the showcase, claiming an 8K image could be sliced into 20 pieces and still show substantial detail .

Generations in 8K resolution with Kling Image 4 Skin texture, loose hair, fabric, film grain. These are the details worth zooming into 4 … Only an 8K image can be sliced into 20 pieces and still show this much detail. 🔍 Love this creative way by [@magnific](https://x.com/magn…
AI High Signal

Andrew Carr argues that “harness engineering as a service” represents a multimillion-dollar annual opportunity for businesses seeking to reduce token use, improve accuracy, and increase repeatability; he predicts harness engineering will matter more than prompting and for longer.

There's millions of dollars a year in \`harness engineering as a service\` for businesses looking to save tokens, increase accuracy, and …
AI High Signal

A proposed data-constrained optimization method, Training-Time Testing (T³), takes a trial step on half a batch, tests it on the other half, and updates with a meta-gradient; its authors claim state-of-the-art results on SlowRun. A commenter characterized the method as “Heun’s method,” a caveat on its novelty.

Compute grows fast but pretraining corpora don’t, so data-constrained optimization matters! We propose Training-Time Testing (T³): each s… This is actually Heun's method! [https://x.com/xingyudang/status/2121024085165531554](https://x.com/xingyudang/status/2121024085165531554)
AI High Signal

Tencent released open weights for Hy4, a natively multimodal MoE with 770B total parameters, 49B active parameters, and a 1M-token context, alongside a system card with a comprehensive safety evaluation report. In a blind evaluation spanning 579 financial tasks and 213 professionals, Hy4 scored 3.04 overall.

vLLM announced day-zero support for Hy4’s text, image, and video capabilities. Its serving path connects the vision encoder and projector to Hy4’s sparse-attention MoE, supports BF16/FP8 on NVIDIA CUDA, and is used in production by inferact.

Today, we’re releasing the open weights for Hy4. Our natively multimodal flagship MoE (770B total / 49B active, 1M context). Alongside th… How does Hy4 handle real work? Hy4 scored 3.04 overall in a blind evaluation spanning 579 financial tasks and 213 professionals. 2/4 ![](… 🎉 Day-0 support for [@TencentHunyuan](https://x.com/TencentHunyuan)’s Hy4 in vLLM! Text, images and video in one model: read visual docum… The serving path connects the vision encoder and projector to Hy4’s sparse-attention MoE, with BF16/FP8 on [@NVIDIA](https://x.com/NVIDIA…
AI High Signal

Kaggle and Eigent AI announced AgenticRL-Bench, which tests coding agents on 22 post-training RL engineering tasks across stacks including Miles, slime, verl, SGLang, Megatron, and Harbor; its leaderboard scores agents on 19 verifiable tasks, each requiring all hidden checks to pass, with three attempts per task. The current leaderboard lists GPT-6.1 Sol at 22.8%, Gemini 4 Carbon at 19.3%, and GPT-6 Astra and Claude Opus 5.5 at 14.0%. Reproducing the benchmark revealed a comparability caveat: seven tasks trigger lab safety classifiers, and trigger rates vary widely across labs, so results may reflect labs’ risk tolerance as well as model capability.

Today, in collaboration with [@Eigent_AI](https://x.com/Eigent_AI), we’re announcing AgenticRL-Bench, a benchmark that evaluates whether …
AI High Signal

A post suggests Anthropic may have found a new gene-editing mechanism, but offers no explanation beyond an attached video, so the claim cannot be assessed from this source.

It looks like Anthropic found a new way of gene editing mechanism. [![Video](https://pbs.twimg.com/amplify_video_thumb/212356057783457382…
AI High Signal

A reported 12-week effort involving nearly 1,000 AI agents across 35 labs produced 936 studies with code and data and an agent-built scientific community with its own peer-review journal; the report says no human wrote the code or text. Agents debated study designs, retracted mistakes, found errors in human papers, and launched a journal that caught other agents cheating in code. In blinded ratings by AI judges, agent papers scored higher on rigor (9.0 vs. 7.0), while human preprints scored higher on novelty.

Super interesting research: almost 1,000 AI agents from 35 labs did research on their own for 12 weeks and ended up building their own sc… Leave one AI agent alone to do research and it often cuts corners. Leave nearly 1,000 AI agents together for three months and they invent…
AI High Signal

An X post congratulated Anthropic on an IPO, describing it as “incredible”; it provides no details or corroboration, so treat the claim as unverified.

Huge congrats to Anthropic on their incredible IPO! I still vividly remember [@8enmann](https://x.com/8enmann) demoing an early version o…
AI High Signal

Tencent Hunyuan says it has shipped Hy4 and Hy4 Flash, describing them as an “ultra-fast, lowest-cost everyday model”; its research team scheduled a Reddit AMA for Nov. 20, 8–10 AM PT, to discuss how they were built.

We’re hosting an AMA on Reddit this Friday, Nov 20, 8–10 AM PT! 🚀 Fresh off shipping Hy4 & Hy4 flash, our research team & [@Shuny…
AI High Signal

MyClaw introduced version 3.0, supporting agents across Claude Code, Codex, Hermes, Pi, and OpenClaw with shared context; it says agents come with built-in tools and access to 5,000+ live data sources. Agents continue running in the cloud after the laptop is closed, with access through browser, Slack, Teams, Telegram, and WhatsApp.

𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐌𝐲𝐂𝐥𝐚𝐰 𝟑.𝟎 Raised AI teammates with tools built in — not a blank agent you babysit. •𝐀𝐠𝐞𝐧𝐭 𝐰𝐢𝐭𝐡 𝐚𝐧𝐲 𝐡𝐚𝐫𝐧𝐞𝐬𝐬 — Claude Code · C… Agents shouldn't be locked into a single harness. MyClaw 3.0 lets you run your agents on Claude Code, Codex, Hermes, Pi, or OpenClaw, wit…
AI High Signal

Theo praised Gemini 4 Neon’s efforts to make it easier to work with, saying it was far ahead, but offered no specific examples or benchmarks.

Whoever lead the "make it easier to work with" efforts for Gemini 4 Neon needs to be put in charge of Deepmind. jfc. They're so far ahead…
AI High Signal
  • Texas elected Gina Hinojosa, who campaigned on a datacenter moratorium, over Greg Abbott; Georgia and Ohio also elected governors who ran on pausing new datacenter construction. Texas environmental permits have been frozen since Sept. 21, and Hinojosa wants the pause to last until new rules pass; depending on when approvals resume, SemiAnalysis estimates the freeze could last about eight months or close to a year.
  • SemiAnalysis says its pre-vote figures show roughly three-quarters of the gigawatts due across its top 20 states in 2027 are already under construction, while Texas accounts for close to 30% of U.S. gigawatts in 2028—making permitting and rulemaking a significant risk to its datacenter forecast.
Before the election, we called a prolonged Texas permit freeze the biggest risk to our US datacenter forecast. Texas then elected the can… Georgia and Ohio also elected governors who ran on pausing new datacenter construction. (2/5) Texas environmental permits have been frozen since Sept. 21. Hinojosa wants a moratorium until the legislature passes new rules, and the … If approvals stay on hold through that session, the freeze lasts about eight months. If reopening waits for new rules to take effect in l… What that changes, using the numbers we published before the vote: 🟠 2027 is largely locked in. Roughly three-quarters of the GW due acro…
AI High Signal

Swyx says an earlier claim that the odds of reaching SOTA had fallen to zero was overtaken roughly three months later, with the effort reaching SOTA . The referenced author says they are on Gemini’s RL scaling team, had printed the article on their desks as a reminder of what had been true, and that “the times” are changing .

when scrolling social media i think it is worth bearing in mind that you can say “odds of reaching SOTA have dropped to zero” and get alo… I want to thanks [@SemiAnalysis_](https://x.com/SemiAnalysis_) for the motivation. In the Gemini RL scaling team, we printed & put th…
AI High Signal

ValsAI’s RSI Index ranks Gemini 4 Neon #4, 3.8 points above Gemini 4 Argon, its previous best Google model; ValsAI says Google’s recent acceleration put it ahead of OpenAI on the index . @scaling01 cautions that the ranking comes from only four tasks run once per model, making scores noisy, and questions whether those tasks represent frontier-lab research; they also warn that benchmark-focused post-training can encourage hill-climbing and distillation. They note the models receive 12–30 hours and suggest broader research, optimization, data, and training tasks would better capture these capabilities .

Gemini 4 Neon is the best Google model on the RSI Index. It lands at [#4](https://x.com/hashtag/4), 3.8 points above Gemini 4 Argon, Goog… I have to say, I find the "RSI Index" name rather irritating and don't put too much weight on it for my considerations of when we get act…