독립 평가, Muse Agent, 반복형 Transformer와 도메인 특화 AI 인프라

Wild 24 hours for AI and lots of different proposals have been made. TLDR; the only *tangible* new fact is that OpenAI and Anthropic are going to have embedded 3rd party evaluators from unknown organizations with Dario floating METR as a possibility. Having 3rd party evaluators is smart as there is no Section 230 style liability shield for model outputs and showing a “duty of care” will be important in future litigation. Several internet companies might have gone bankrupt without Section 230 so limiting liability really matters. There are minimal investment implications from this single new fact, but I do think that for anyone who wants a “smoother for longer” cycle then most constraints are good: wafers, watts, real rates and spreads. Excessive regulation is a different matter but I don’t think we are anywhere close to this even if the vector changed over the last 24 hours. To summarize the events: Dario made the most maximalist proposal of the weekend: embedded 3rd party evaluators, a national regulatory regime for models beyond a certain capability/ingredient threshold, a broad international regulatory pact between democracies, stricter limits on compute/distillation for China and then a different international regulatory regime that encompasses China. Before there is a national regulatory regime, he wants a Sherman act waiver so that Anthropic can safely coordinate with OpenAI and other frontier labs without antitrust fears. TBF, this latest proposal is much less maximalist than some of his prior proposals like “Policy on the AI Exponential,” where he advocated for an FAA for AI. I believe he is sincere in his beliefs. And despite all the protestations, all of this would also probably be good for his business over the long-term. Sam agreed that embedded 3rd party evaluators were a good idea and stated they would implement them. Again, this is smart as should help limit future liability. Elon said “Dario is right” and later specified that “Dario is right that there should be some oversight. Peer review of AI by competitors is the right way to start this off.” This would be a MPAA like self-regulatory structure for AI with regular calls between the labs plus a process where each new model is evaluated for safety by competitors for a 1-2 week period before being released. That is *wildly* different from Dario’s proposal and in-line with what David Sacks has been proposing. Elon also stated that nothing was going to slow down open-weight models. Demis said that Dario’s essay was a “step in the right direction.” Dario also said that he was also open to Demis’ idea of a FINRA like self-regulatory structure as part of his proposal. David Sacks had a thoughtful post where he said that Dario and Sam should pace unilaterally, called the antitrust waiver a cartel request and denied that METR was truly independent given their ties to Anthropic. Sriram Krishnan, former White House AI advisor, noted that it would be important to have the 3rd party evaluators come from independent organizations that are not affiliated with any lab, which is basically an indirect statement about the relationship between METR and Anthropic which Sacks was explicit about. Clem from Hugging Face said they were open to being a neutral 3rd party evaluator, which is interesting especially if Jensen was consulted before that post. Alexander Wang from Meta noted that alignment would be an increasing focus going forward. An executive order seems likely after all this and the language in this EO is going to be really important. It is possible to democratize and distribute AI broadly and safely without centralizing it in the hands of a few corporations who might each become more powerful than any single government. I do not want a few humans in control of intelligence. I want us all to have our own intelligences that reflect our own values and human variation in all of its richness. Intelligence distribution over intelligence centralization FTW.

surely the us government isn't about to let frontier labs select their own former coworkers, friends, sex partners and equity holders to be the new regulators and gatekeepers of ai

Looped transformers are a popular architecture topic right now. This new technical report extends the loop across tokens. Recurrent Looped Transformer (RLT) makes the decoder recurrent over every token, prompt and response included. A causal encoder builds global KV memory. For each new token, the decoder combines the token's encoder representation with its own final hidden state from the previous token and a sliding-window cache of recent activations. With a 48-layer decoder, the computation path after t tokens runs through 48t decoder blocks, while each token still executes a fixed number of blocks. Depth grows with the sequence and per-token cost stays the same. The same state transition is used for pretraining, SFT, sampling and RL replay, and nothing resets at the prompt-response boundary. RL replay rebuilds states under the current weights instead of reusing stale rollout states. The report is a design proposal. The author states that reasoning gains, hardware speedups and RL scaling are goals that have not been measured yet. Paper: https:// github.com/yifanzhang-pro /recurrent-looped-tranformer … Chat with Paper: https:// academy.dair.ai/papers/recurre nt-looped-transformer …

Alexandr Wang

@alexandr_wang

2일 전

a big part of the team’s focus when building muse was to make something that “just worked”. so much work went into all the little details to make the product simple and easy-to-use. been awesome to see so many people from different walks of life find muse useful!

Alexandr Wang

@alexandr_wang

2일 전

muse code + muse spark 1.3 max is really good at real-world software engineering

Jerry Liu

@jerryjliu0

2일 전

in a world where frontier models can generally do everything, it makes sense that every startup AI co is a "domain specific harness" * at an extremely basic level it helps for branding and distribution. you show that you care a lot more about a specific domain than the labs * re: tech you can define more domain-specific evals and hillclimb towards it with higher accuracy, lower cost, and lower latency vs. using the models * re: product you can define a more guided experience that's more intuitive to execute a workflow, store and manage data, etc.

Shreyansh Patni

@shreyanshpatni_

2일 전

Want to build a domain-specific agent harness? This is a great practical guide by @sydneyrunkle on designing and building your own custom agent harness. Worth reading if you’re exploring agentic systems and AI-native products. https:// langchain.com/blog/how-to-bu ild-a-custom-agent-harness …

Jerry Liu

@jerryjliu0

2일 전

Introducing FlyOCR - I trained a fly brain to read a PDF It uses the full MaleCNS v1.0 fruit fly connectcome. The architecture is inspired by doomfly by @nftechie_ The fly splits a pdf image into individual glyphs, maps pixels into receptor activations, runs simplified current-based dynamics across the 166k neurons and 25m edges in the circuit, applies a compact readout model on the downstream spikes, and concatenates everything into the parsed output. On reading an actual Microsoft 10-k, the fly gets ~86% over the balance sheet heading, but is largely able to read the numeric values correctly. Over 1.7k+ sampled glyphs (chars+digits) it gets 87% accuracy. With enough training it might match some of the latter-generation MNIST models! Maybe eventually we’ll replace our doc parsing VLMs with flies. Full video below. Repo with full code + report: https:// github.com/jerryjliu/fly_ ocr …

Jerry Liu

@jerryjliu0

2일 전

if u don't want to use a fly brain to read a PDF, you could always check out an actual OCR solution like llamaparse https:// cloud.llamaindex.ai

“SenseNova-U1.5: Towards Native Unified Visual Intelligence” Most multimodal models still use separate representations for seeing and generating images. SenseNova-U1.5 instead uses one 8B encoder-free, VAE-free model to understand, reason about, generate, and edit images directly in pixel space, including native 4K generation. It then trains specialist RL experts for aesthetics, text, infographics, and editing, and distills them back into one unified model. The result is a stronger case that visual understanding and generation can share the same native representation rather than being separate systems. https:// alphaxiv.org/abs/2609.11929

“BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure” Reward hacking in agent benchmarks is often an infrastructure problem, not just a model-behavior problem. So this paper formalizes the full reward path and instruments runs to distinguish vulnerable tasks from actual exploit use, reaching 96% runtime detection accuracy and much higher exploit-chain recall than prior scanning baselines. https:// alphaxiv.org/abs/2609.11028

You can't call AI videos slop anymore. 100% AI