We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Top Stories
Why it matters: The frontier contest is shifting from release cadence to control over training, evaluation, and deployment.
Pacing moved inside the training loop. Anthropic says it will give third-party evaluators permanent, employee-level access to verify safety measures, report incidents, and assess alignment during training. OpenAI says it now prepares explicit safety cases before frontier reinforcement-learning runs expected to significantly increase capability, seeks shared standards for misalignment and monitoring, and defines pacing as slower progress—not stopping. Sam Altman frames the two failures to avoid as loss of control and excessive concentration of power. The implementation fight is immediate: Cohere CEO Aidan Gomez calls for “evidenced standards, not a cartel,” while @suchenzang assigns near-zero probability to a single evaluator that is competent, financially independent, and willing to speak up.
The enterprise bottleneck is data, not only model quality. The Turing Post says 79% of enterprises are building agents but only 11% have reached production; it points to incomplete or outdated corporate data and describes a memory-plus-retrieval layer as the practical fix.
Research & Innovation
Why it matters: The strongest technical signals are adding recurrence, explicit procedures, and parallelism rather than simply scaling model size.
Recurrent Looped Transformer makes depth recurrent. The proposal applies a recurrent decoder across every prompt and response token. In a 48-layer design, the path grows to 48t decoder blocks after t tokens while each token still executes a fixed number of blocks; the same transition spans pretraining, supervised fine-tuning, sampling, and RL replay. It remains a design proposal: reasoning gains and hardware speedups have not been measured.
Long-horizon agents are becoming inspectable. Google’s Procedural Graphs encode “what to do next” as procedure-to-procedure relations and keep graph edits only when held-out validation holds or improves. PARSER parallelizes document reading and then iteratively queries the evidence; the roundup reports 4B beating sequential memory by 5.7 points on average and 12 points at 896K tokens, while 9B beats DeepSeek-V4-Pro by 6.3 points and cuts latency up to 11×.
Products & Launches
Why it matters: Agent competition is moving toward secure execution, browser handoff, and multimodal workflows that ordinary users can adopt.
Meta’s Muse packages a personal agent as a product. It runs tasks in a secure VM with Sentinel monitoring, offers 100 million free weekly tokens, and uses single-use Stripe payment cards. Meta says Spark 1–1.3 were built specifically for Muse; one user report describes fast browser use with takeover and puts Spark 1.3 at $1.50 per million input tokens versus $5 for Opus.
MiniMax H3 pushes open video toward production speed. The open-weight model offers native stereo audio and multimodal reference control. Its ecosystem includes four-step distillation and NVIDIA’s report of 15 seconds of 768p video with audio in 6.6 seconds of warm inference on eight B300s, excluding loading, compilation, and encoding.
Industry Moves
Why it matters: Capital and compute commitments are consolidating around frontier infrastructure and sovereign AI capacity.
- Mistral financing: A weekly digest reports a €3 billion Samsung-led Series D at a valuation above €21 billion, calling it the largest equity round for a European technology company.
- Anthropic compute: A post reports that Rum, the neocloud formed after Rumble acquired Northern Data, signed a $13.7 billion agreement with Anthropic; Anthropic also receives a warrant tied to future compute purchases.
- Sovereign evaluation: Artificial Analysis signed an MOU to become an official independent evaluator for South Korea’s Sovereign AI Model Foundation; its benchmark already represented 25% of the latest competition’s judging criteria.
Policy & Regulation
Why it matters: US policy is not converging on the labs’ self-imposed pacing plan.
A FT-linked feed report says President Trump rejected calls for an AI slowdown. Altman welcomes a federal framework but says labs should act before legislation; Lina Khan argues existing consumer-protection and FTC laws already reach dangerous, unvetted, or defective AI systems and should be enforced alongside any new regime.
Quick Takes
Why it matters: AI’s externalities are appearing in communication, interfaces, and the machinery used to monitor agents.
- Cold outreach: Random Walker reports about 75 inquiries ahead of the Fall 2027 PhD cycle and says AI-generated signals of interest have made the cold-email channel “effectively dead.”
- Telephony: OpenAI’s 1-800-ChatGPT service is now powered by the GPT-Live SIP telephony API.
- Monitoring: A proposed policy target is a minimum monitoring-compute/inference-compute ratio—preferably 1—while the technical problem of preventing monitor models from becoming sympathetic to what they assess remains unresolved.
- The discussion characterizes the Collatz Conjecture as more of a literature-review and tedious-grind problem than one with no plausible starting point. Astra reportedly finds relevant literature, often directly about Collatz, and can make partial progress on the tricky parts, although each partial result creates another obstacle.
- The participants speculate that an LLM with enough tokens and patience could navigate the conjecture’s many grind-y blind alleys, potentially beyond practical human effort, but say the problem likely requires more theory-building than the system can currently do.
OpenAI says it is “categorically” impossible that a mathematician’s user data from the last two months influenced its “monumental solution” to a Millennium Prize problem; mathematician Tristan Buckmaster said, “I just don’t believe them.” A separate commentary cautions that ideas may emerge independently from shared knowledge and that claims an idea was “stolen” require tangible evidence.
- MiniMax H3 is presented as an open-weight video-generation model with native stereo audio and multimodal reference control. Its ecosystem includes 4-step FastH3 distillation, NVIDIA SANA’s two-stage H3 + LTX-2.5 pipeline, VDN’s released weights/training/inference code, PDD-based 8-step Acc-LoRAs in ComfyUI, and 4-/8-step LightX2V Turbo LoRAs for text, image, and reference-conditioned video-and-audio workflows.
- NVIDIA’s SANA team reports that, on 8×B300 GPUs, its pipeline produces 15 seconds of 768p video with audio in 6.6 seconds of warm inference; the timing excludes model loading, compilation, and MP4 encoding.
- One AI-safety analysis argues that pausing algorithmic progress while allowing compute capacity to keep expanding could reduce near-term accidents but create greater long-term risk: when progress resumes, accumulated hardware could enable an unexpected, uncontrollable “FOOM”; it therefore recommends roughly the opposite approach for minimizing risk.
- The analysis also says safety strategies can be paradoxical: the Hugging Face incident, which better preparation might have prevented, was nevertheless the most safety-promoting event of the year so far, making the ultimate effects of interventions difficult to predict.
Rumble reportedly became a neocloud after acquiring Germany’s Northern Data and later relaunched as Rum; it then signed a reported $13.7 billion agreement with Anthropic, which also received a warrant to buy Rum shares tied to the amount of computing power it eventually purchases.
Anthropic is unilaterally committing to a safety-governance measure in Dario Amodei’s “We Must Pace the Frontier” essay: it will give third-party evaluators permanent, employee-level access to its systems to verify safety measures, report incidents, and assess model alignment during training.
- OpenAI says it now prepares explicit safety cases before frontier reinforcement-learning runs expected to significantly increase capabilities, extending safety work beyond evaluating completed models to the development process itself.
- @sama identifies two major AI risks: losing human control to AI and excessive concentration of power in one person, company, lab, or country; he argues for a middle path that preserves human control while avoiding dangerous centralization.
- The post supports consistent federal safety requirements for frontier AI, including ideas such as independent auditors, and calls for industry-wide standards covering misalignment, monitoring, and safety. It frames “pacing” as slowing progress enough to fund safeguards—not stopping progress.
Joshua Saxe warns that recent agent-hacking demonstrations indicate models can execute complete cyber kill chains autonomously, dynamically discover zero-day vulnerabilities, and overrun companies including OpenAI and Hugging Face—suggesting AI could affect offensive cybersecurity more than incrementally. He projects rapid escalation as systems that currently require eight H100 GPUs potentially run on one within a year or two, while faster token generation, improved hardware, and public-cloud access enable attackers to deploy agent swarms. Saxe contrasts traditional, signature-detectable worms with a hypothetical adaptive, deceptive, mutating, self-replicating agent swarm that could cause a new level of harm.
- QuixiAI claims it trained Andrej Karpathy’s Shakespeare model on a fruit-fly brain connectome.
- QuixiAI introduced MaleCNS, a fruit-fly connectome from HHMI Janelia and Google Research distributed in a Hugging Face-compatible format with original neuron IDs, connectivity, and synapse counts; related GitHub and Hugging Face resources were shared.
- OpenAI now prepares explicit safety cases before frontier reinforcement-learning runs expected to significantly increase capability, extending safety work beyond completed-model deployment to the development and evaluation process.
- Sam Altman supports a federal framework with consistent frontier-AI safety requirements and independent auditors, while urging labs to act before legislation; he defines “pacing” as slowing progress enough for safety cases and monitoring—not stopping it—and calls for shared industry standards and international government coordination.
Muse agent and Muse Spark 1.3 — user-reported strengths and price: Savar Sareen says Muse is his default personal agent and Spark 1.3 his coding model, praising near-instant responses, browser actions with user takeover, context-driven goal and idea suggestions, and a balance of intelligence and directness in coding and non-coding tasks. He reports Spark 1.3 at $1.50 per million input tokens versus $5 per million for Opus on OpenRouter, but prefers Gemini/Nano Banana for image generation; he explicitly frames this as personal opinion, not a statement as a Meta employee.
- Muse Spark 1 through 1.3 were deliberately developed for Muse, with successive releases focused on improving agentic and multimodal capabilities as groundwork for Muse’s personal-agent product. The team says its longer-term objective is “personal superintelligence,” with Muse positioned as its practical manifestation after more than a year of coordinated model, infrastructure, and data work.
Muse says it developed the Muse Spark 1–1.3 models specifically to power its Muse personal agent, with successive releases focused on major gains in agentic and multimodal capabilities. The company frames Muse as the culmination of more than a year of work toward broadly available “personal superintelligence.”
Anthropic is unilaterally committing to give third-party evaluators permanent, employee-level access to its AI systems so they can verify safety measures, report incidents, and assess model alignment during training. The commitment is presented as the first step in Dario Amodei’s proposed three-part plan for slowing the AI industry.
T3 Code now streams responses by default, but not token by token. The streaming behavior is planned to be toggleable before it reaches stable.
- V4.1 Pro was described as strong at Python in an informal out-of-distribution test involving programmatic recreation of a complex surrealist image from a text prompt, without image reuse or GUI stroke automation.
- The commentator speculated that V4.1 might score 75–78% on ARC-AGI-2; this is an estimate, not a reported benchmark result.
Astra was reportedly able to animate a full sequence from scratch in an early demonstration; the author acknowledges that the result “isn’t good” and speculates that a sufficiently long reinforcement-learning environment could eventually let it “move the world.”
METR is the subject of an independence debate. Neel Nanda says METR publicly lists its funders and does not accept funding from potentially compromising sources such as AI labs and “coefficient giving”; he considers its finances acceptable while acknowledging other independence critiques. He characterizes company-provided free tokens as compensation or expense coverage rather than funding and highlights the diversity of METR’s funders.
Muse is being praised for thoughtful product design that onboards mainstream users who are not engaged with AI Twitter. The team says it is refining every detail to make Muse easy for everyone to use.
@scaling01 gave Astra a skeptical hands-on assessment: after using it exclusively and burning through two Pro limits, they said it did not feel like AGI, criticizing its lack of judgment or “taste” and tendency to act without direction, wasting time and tokens; they nevertheless called it the best model available, albeit essentially a somewhat larger “code monkey.” The commentator also said Astra was “much smaller” than expected and speculated—without presenting verification—that it may be 6–8T parameters rather than the rumored 10T+.
