We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Frontier-level cyber offense is now available as open weights
Vals AI released CyberOffense. In the benchmark, each agent works alone from an isolated attack machine against two simulated enterprise networks. It gets a public web address and no human help. Vals says the results show agents can now break into realistic corporate networks unaided, and that this capability is "available to anyone" .
- Scores. GPT-6 Sol scored best at 80.8% of objectives. It is only available through OpenAI's restricted Daybreak cyber program. DeepSeek V4.1 Flash 1017, an open-weight model, scored 72.8% and matched Daybreak on the Nova network. On Lighthouse, Daybreak completed 27 of 31 objectives in under three hours, and DeepSeek completed 21 .
- Cost and token budget. A DeepSeek run cost under $3, against about $22 for Daybreak, and the open model ran with no safeguards. DeepSeek completed nothing in its first 25M tokens and made most of its progress after 100M. Vals concludes that short tests understate what open models can do. It also expects attacks like these on real networks to become more common in the coming months, and urges defenders to use the same models to find and fix weaknesses first .
Vals' point is that gated access matters less once open weights are this close behind. OpenAI's restricted Daybreak program, covered in recent briefs, is the clearest example. Google is following the same gated approach: per a newsletter summary, Gemini 4 Carbon tops the Artificial Analysis Intelligence Index at 59, largely on cybersecurity gains, but only vetted cyber defenders can use it .
DeepMind cryptanalysis. Google DeepMind reports that AI agents, working under human supervision, found two weaknesses in legacy RSA-1024:
- Signatures. Padding adds no protection once an attacker controls roughly half or more of the signed block. For cases where the attacker controls less, a new algorithm is estimated to improve on the previous best method by about 1,600× .
- Batch factoring. A new algorithm cuts the average cost of factoring each additional 1024-bit key by about 30× .
DeepMind stresses that modern 2048-bit and larger keys remain secure. Its argument is that AI can now probe hardness assumptions much faster than the years or decades such advances used to take .
Inkling 2 Small: strong ARC results and signs of a cheap training run
Thinking Machines launched Inkling 2 Small last period. This period brought third-party results:
- ARC-AGI. ARC Prize verified 81.7% on ARC-AGI-2 at $0.27 per task and 96.5% on ARC-AGI-1 at $0.08. It says the model beats every open-weight model it has evaluated on both .
- Code Arena. Arena lists it at 1648 points, #8 among labs in Code Arena: WebDev. It would be the #1 US open model once weights ship .
The compute behind it is drawing attention. One analyst estimates pretraining used a bit under 1,000 GB300 GPUs and about 130K GPU-hours, roughly $650K at rental rates. That is his estimate, not a disclosed figure .
Funding. A VC post describes a "$5B @ $40B" deal. It says Thinking Machines has committed to deploy more than 1 GW of Vera Rubin compute, which the post puts at $50B+, and has over $100M in ARR .
Behavior under evaluation. Vals found profanity in some of the model's benchmark runs. It was not a reliable sign of failure: 96% of failed runs never swore, and two looping runs produced 241 of the 316 uses .
AI-assisted science: a math proof and two new lab benchmarks
- Aaronson–Ambainis conjecture. Physical Superintelligence says its research system, using a non-public Anthropic model, proved this conjecture in quantum complexity, open since 2008. The system reportedly found a complete proof on its third day, and the final version was checked in Lean .
-
VIALS-AA. Artificial Analysis launched this leaderboard of 161 visual-reasoning tasks drawn from life sciences, such as gel blots, flow cytometry and plasmid maps. A task counts only if the model answers correctly in all three runs.
- Gemini 4 Carbon leads at 64%, followed by GPT-6 Astra at 63% and Claude Opus 5.5 at 61%. Kimi K3.1 Pro is the best open model at 45%.
- Cost per task varies widely. Opus averages $1.64, about 29× Sol's ~$0.056, largely because it takes about 27 turns per task against about 8 for Astra and Sol .
- BioBasicsBench. OpenAI and Ligo released 100 tasks taken from real experiments. They test whether agents can interpret raw lab data, catch errors, and recognize when evidence does not support a conclusion. OpenAI frames the benchmark partly as tracking biosafety and biosecurity .
China, compute and the Nvidia–Megaspeed reporting
"Distillation lag." An X article translates minutes of a private October discussion among Chinese AI insiders. They argue the commonly cited "3–4 months behind" reflected the time Chinese labs needed to distill US models, not underlying capability. Their claims:
- Distillation costs have risen 5–10×, and up to 20× in some cases, as US labs ban suspected distillers and hide reasoning traces .
- Chinese labs hold the equivalent of 10K to a few tens of thousands of Blackwells, against a projected 2–3 million for a single US lab next year. On that basis the room expects the gap to widen to about a year, or 18–24 months if US labs reach recursive self-improvement .
- Observer teortaxesTex calls the distillation thesis "largely vindicated" but doubts the claim that Beijing would back DeepSeek .
Huawei capacity. Epoch estimates Huawei's Horinger data center at about 80K H100-equivalents, or 7% of Colossus 2's ~1.1 million .
Megaspeed. Commentator Chris McGuire posted a recap of FT, NYT and Bloomberg reporting:
- US intelligence assessed that Megaspeed executive Alice Huang was a Chinese spy, and that her goal was to get Jensen Huang to lobby for looser export controls.
- Megaspeed is likely a Chinese cutout and remains a preferred Nvidia cloud partner receiving Blackwell chips.
- BIS has not tightened controls on Megaspeed or Southeast Asia .
These are the recapper's framing of others' reporting. They add pressure to the export-control debate at the same moment Chinese insiders describe compute as the binding constraint.
Baidu. Baidu released open weights for ERNIE-5.5-Flash-Base, a 280B-parameter MoE (14B active) trained on 33.5T tokens. In Baidu's preliminary evaluations it beats DeepSeek-V4-Flash-Base, scoring 78.4 vs 69.2 on MMLU-Pro. Its license excludes the EU, UK and South Korea .
Products and infrastructure
- Meta. Muse and Meta AI now run specialized models trained with Muse Wonder. They add "Active Memory" and "Active Focus," and Meta claims a 2.4× reduction in forgetting goal-critical information and 67% better recovery when constraints change .
- Anthropic. An Engineering Blog post describes "dreaming," a process by which Claude organizes and corrects what it has learned so it can find it more easily later .
- GPT Image 2.6. The models take #1 and #2 on Artificial Analysis' text-to-image and image-editing leaderboards. Flare is almost twice as fast as 2.5, and pricing is unchanged at about $0.21 per max-quality image .
- Vera Rubin inference. On agentic workloads with DeepSeek V4.1 Flash, NVIDIA's Vera Rubin configurations deliver 5.0× GB300's throughput per GPU at 260 tok/s per agent, but only 1.7× at 80 tok/s per agent .
- GPU market. SemiAnalysis estimates at most 2–3 GW of the ~50 GW of capacity due online in 2027 is at risk. One-year H100 contract prices are up about 21% since March .
- Anthropic IPO. Per the WSJ as relayed on X, Anthropic's annualized revenue run rate topped $80B in October. Investors are questioning a $1.5T IPO valuation .
Watch
METR follow-ups. Ajeya Cotra says the METR authors' estimate that AI nearly doubles capability progress amounts to researchers experiencing time about 10× faster. A naive extrapolation implies ASI within a year, if bottlenecks don't bind . Ryan Greenblatt says he trusts the ~2× estimate but finds it reasonable to want the redacted evidence .
Releases still pending. Qwen 4 was expected this week based on remarks by Alibaba Cloud's CTO . No new frontier model from OpenAI or Anthropic shipped by period's end .
@kimmonismus said the U.S. midterm elections seemed detrimental to U.S. AI development and linked a post that reads “Ban RSI.” The posts do not explain “RSI” or provide specifics for the election assessment.
A post advocates “human disempowerment” and an end to human politics, arguing that the median AI exceeds the median human voter in knowledge and decency. @inerati rejects that position and says its proponents should be kept far from the levers of power.
Rep. Casar posted a terse call to “Ban RSI.” @vikhyatk replied that doing so would require AI development to be done by humans, joking that he supports the policy and misses manually debugging PyTorch shape errors.
@teortaxesTex argues that reaching the Singularity will require larger models, describes Opus 5.5 High Ultracode as poor when left to its own devices, and considers publicly available AGI easy to criticize; these are the author’s judgments. In a follow-up linking that post, they say “OK Opus, this is more like it” and attach a video, without explaining in text what prompted the reaction.
An in-app browser is slated to inherit users’ cookies and passwords when they opt in, matching the agent’s own browser tool; the change is planned for the next patch release.
Posts position Thinky Machines’ Inkling 2 Small as a near-frontier open-weights offering that can be fine-tuned on-premises or through Tinker and is aimed at sovereign use cases; a related post describes Inkling 2 as the best open-weights model outside China (the author’s assessment) and reports Thinky has $100M+ ARR. The posts report that Thinky committed to deploy more than 1 GW of Vera Rubin compute, with the quoted post estimating the compute cost at $50B+. A quoted VC post labels a Thinky Machines deal “$5B @ $40B,” but the excerpt does not specify its terms.
@theo says “T3 Code already won” in an exchange about building an ADE. The quoted post calls ADE-building a “tarpit,” claiming there are 500+ such tools, questioning whether Conductor could have been a long-lasting business, and arguing that B2C developer products are hard to monetize because developers do not want to pay; these are the posters’ market assessments.
David Kasten argued that Jensen Huang might be a victim rather than a knowing agent of a Chinese intelligence officer, but nevertheless called him “compromised” and said he should not remain in his role, linking an FT article.
A post’s recap of FT, NYT and Bloomberg reporting says U.S. intelligence assessed that Megaspeed was likely a Chinese cutout helping Chinese firms obtain AI chips or circumvent export controls, and that executive Alice Huang sought to persuade Nvidia CEO Jensen Huang to lobby for looser restrictions . The recap says that, despite intelligence officials briefing senior U.S. officials, BIS allowed H200 exports to China and did not enforce licensing requirements for advanced Blackwell shipments to Megaspeed and Southeast Asia; Megaspeed continued receiving Blackwell chips as a preferred Nvidia cloud partner . It also claims tens of billions of dollars in AI chips were exported to Southeast Asia that year, with Malaysia and Singapore among the world’s four largest AI-chip importers .
A post proposes investigating whether AI models can be linked to code that later receives CVEs, then using model-specific vulnerability rates to set a minimum security bar for models used in foundational software; it warns that without such a bar, competition may favor cheaper tokens.
Research on subliminal learning finds that regularization amplifies the transfer of hidden biases; the researchers suggest these biases may be simple, natural explanations of the data, while more complex traits such as backdoors appear to favor less regularization. A commentator argues the result should motivate less distillation, raising concern that simple biases could transfer even if backdoors are less robust to regularization.
The post describes Inkling’s pretraining run as comparable to DeepSeek-Coder-V2 (June 2024), reporting 21B active parameters and 10.2T tokens for Inkling versus 17B and 12.25T for V2. It says Inkling 2S was substantially cheaper and finished 12× faster, while noting that much of the cost difference reflects hardware and precision differences (BF16 vs. FP4).
Togelius argues that frontier models remain weak at investing and can be beaten on specific investment tasks, while some trading styles and financial tasks are easier to train for. He frames Buffett- and Munger-style investing as especially challenging for RL because it involves long horizons, very sparse rewards and actions, and highly imbalanced rewards. He adds that broad world knowledge does not guarantee understanding or knowing which knowledge to apply where.
- An analysis of a private discussion among Chinese AI insiders argues that the commonly cited 3–4-month China–US frontier gap reflected distillation lag rather than capability. Participants say suspected-distiller blocks, degraded responses and restricted reasoning traces have made distillation costlier; they estimate costs rose 5–10×, up to 20× in some cases. Coding remains distillable at rising cost, but computer-use, spatial and physical-world models are harder to distill.
- As distillation narrows, compute becomes a central constraint: the article estimates Chinese labs have the equivalent of 10,000 to a few tens of thousands of Blackwell GPUs, compared with a single US lab potentially having 2–3 million next year. It relays a forecast that the gap could widen to about a year—or 18–24 months if recursive self-improvement occurs; these figures and forecasts are not verified outcomes.
- Cheap US models threaten Chinese models’ low-cost differentiation: the article reports GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens, with an index score of 28, versus 28 for GLM-5.3-Flash and 33 for DeepSeek V4.1 Flash; it argues falling US prices could pressure Chinese model economics. The article sees compute-owning hyperscalers and possibly DeepSeek as better positioned than independent labs, but the poster disputes the claim that Beijing would back DeepSeek, while agreeing that the distillation thesis is largely vindicated.
QuixiAI recommends lazy-hq/aisdk to developers building agents in any language and argues that Vercel should own it.
Swyx committed to weekly coverage of Anthropic’s potential IPO until it happens or does not. An associated discussion examines Anthropic’s S-1 for lessons on the relative economics of “neolabs” and “neoclouds” and the durability of Anthropic’s competitive edge. Whether Anthropic will IPO—and whether it could be valued above $1.5T this year—are posed as open questions, not reported outcomes.
@TheWholeDaisy says some leading agentic coders can regain a flow state by switching among enough agents without waiting, but only when they can sustain “30s per agent for hours.” The post they quote describes agents as impressive but still frustrating, with the field early and a long way to go.
@teortaxesTex estimates that Thinking Machines used about 13–14 NVL72 GB300 racks (slightly under 1,000 GPUs) for pretraining, followed by RL on 2,000 GPUs; the post puts pretraining at 130,000 GPU-hours and about $650,000 at rental rates, and says it was faster than DeepSeek-V2’s run through 1 trillion tokens on 2,000 H800s. These are the poster’s estimates, not confirmed figures.
A post showcases Inkling 2 Small as able to paint and labels its video and audio as made in code; it also presents a same-prompt comparison of GLM 5.3 Flash, DeepSeek V4.1 Flash, and Qwen 3.8 Flash, without stating comparative results in the text.
A developer demonstrated Hermes Agent controlled from a Nintendo 3DS as a remote for a Mac, where the agent can search the web and write and run code. Voice input uses a custom 3DS homebrew app because the browser cannot access the microphone; audio is sent to the Mac for Whisper transcription, and the 3DS presents approval requests with A/B controls to approve or deny.
