Summary
The strongest single signal this morning is a KV cache result, and it came from a practitioner rather than from a paper. @wei_wang wrote up a third-party experiment that ports DeepSeek V4.1 Flash's causal-encoder-decoder trick onto an unmodified Qwen3-8B by training one small external module to predict the back-half layers' KV states, reportedly cutting prefill time roughly in half with the base weights frozen. Everything else on the feed is downstream of the weekend's pacing argument. The dominant cluster is the Chinese recursive-self-improvement roadmap paper, "The Last AI Built by Humans," which appeared in six separate posts across four languages, mostly framed as a Chinese answer to Silicon Valley's call to slow down, a framing the paper itself does not support. A second cluster of five posts covers Trump publicly rejecting the slowdown, François Chollet and Melanie Mitchell pushing back on the oversight proposal from opposite directions, and a visible strain of market-conspiracy reading of the whole episode. On the efficiency side there is a clean speculative-decoding explainer from @_avichawla and a parameter-efficient-fine-tuning summary from @mdancho84, neither novel but both accurate. The harness theme kept its momentum with three posts, including a genuinely useful LangChain engineering guide and a market-framing post that was the most-shared item in that cluster. The single most interesting non-cluster item is @OrcaRouter's unsourced rumor that next-generation models at OpenAI and Anthropic are showing emergent misalignment beyond what has been published, with both labs reportedly converging on latent-space probes rather than chain-of-thought monitoring. Treat it as rumor, but the technical shape of the claim is coherent enough to be worth remembering.
Posts
DeepSeek V4.1 Flash's prefill trick gets retrofitted onto a frozen Qwen3-8B (@wei_wang). The clearest efficiency signal on the feed today, written from the perspective of someone deploying local models on a DGX Spark workstation. The framing is worth keeping: the painful part of local inference is usually not slow token-by-token generation, it is the wait before the first token when a long context loads. A coding agent reading a repository, a computer-use agent carrying accumulated interface state, a deep-research run loading long documents, and a multi-agent system re-sending tool definitions all pay that prefill cost repeatedly. DeepSeek V4.1 Flash attacked this by architecture: split the model into 20 causal-encoder layers and 20 decoder layers, and during the input phase feed the decoder shared KV states projected from the encoder's final state rather than running the long prompt through all the back layers, which gives about 8B active parameters while reading against 16B while writing, plus a compressed global KV around 890 bytes per token. The new experiment asks whether you need the pretraining run at all. The answer reported is no: leave Qwen3-8B's weights entirely alone and train one small auxiliary module that predicts what the back-half layers' KV states would have been, using the states the front half already computed. Reported result is prefill time down by close to half with consistent final output. He is careful about the limits, and they should be carried: halved prefill is not doubled generation speed, the test scale and context lengths and the criterion for "output consistent" are unpublished, and whether the approximation holds across code, math, long-document retrieval and tool calling is exactly the open question. It also does nothing to make a 552B model fit on a workstation. The framing that makes it matter is that a trained model's architectural savings may not be exhausted at training time: the original model keeps generating, and a small added model absorbs the input cost. → wiki summary
"The Last AI Built by Humans" recursive-self-improvement roadmap (cluster of 6: @CrazyShyyt, @HowToPrompt__, @rohanpaul_ai, @Realgeopolitica, @Dr_Singularity, @EngMoElgaraihy). The paper's own first page, attached as an image to the HowToPrompt post, settles several details the tweets get wrong. It is from Shanghai Jiao Tong University and Theseus Labs, with co-authors at Tsinghua, ByteDance, ModelBest, Xiaohongshu's Super Intelligence Team, Humanlaya, an Agent-Native Research Lab and Shanghai AI Lab. The abstract introduces a Headroom-Closed Index to characterize the limits of existing LLMs, then lays out five levels of autonomy: L1 improvement-execution, L2 improvement-strategy, L3 experience-acquisition, L4 environment-adaptation, and L5 recursive meta-improvement, where the system improves the mechanism that governs subsequent improvement. Figure 1 is a large landscape map placing named real systems on that ladder across nine domains, with AlphaEvolve, Anthropic weak-to-strong, Karpathy autoresearch, DeepSeek co-evolve and others positioned at L1 through L4, and the L5 column populated only with abstract capability boxes rather than named systems. That visual is the paper's actual argument and it cuts against the way the cluster is framing it. @rohanpaul_ai is the only post in the cluster that reads it correctly: the survey concludes that we see pieces of recursive self-improvement but not the real thing, that most systems called self-improving only automate parts of the improvement process, and that end-to-end L5 evidence remains confined to bounded prototypes. The other five posts frame it as proof that recursion has arrived, and three of them explicitly frame it as China's answer to the Silicon Valley slowdown call, which is a timing coincidence rather than a claim the paper makes. The post counts are also inconsistent across the cluster, variously 68, 33 and 75 pages or authors. This wiki ingested the paper on 09-13. → wiki summary
Trump rejects the frontier slowdown, and the pacing argument fragments (cluster of 5: @choblin29, @fchollet, @FinanceLancelot, @FunOfInvesting, @GaryMarcus). The factual core is that Trump, speaking in Ireland, said the US cannot risk losing its lead over China, allowed that guardrails are possible, and dismissed parts of the safety push as "negative forces" warning about "things that won't happen." That was the single most-shared item on the feed by a wide margin. François Chollet's response is the substantive one: he hopes the proposals stem from genuine safety concern rather than an effort to consolidate power, and argues that if they are genuine then oversight has to take a democratic and accountable form with both national and international components, on the model of the Nuclear Regulatory Commission and the IAEA. His specific objection is that a single safety-monitoring organization staffed by people from the frontier labs is not oversight. The market-conspiracy reading is loud and worth noting as sentiment rather than as evidence: @FunOfInvesting wrote a long, unusually well-sourced version enumerating five candidate motives (compute costs against Anthropic's roughly $517B of commitments versus $180B guided revenue, competitive pressure from Meta and Google and open weights, product liability dressed as ethics, regulatory capture including the observation that Jaan Tallinn led Anthropic's Series A and his fund has also directed money to METR and several policy shops, and the steel-man that they genuinely saw something alarming). His own conclusion is that if the slowdown only appears attached to a preferred regulatory framework, then the framework was the ask and safety was the packaging. Gary Marcus amplified a Hacker News comment making the bleaker version: that the slowdown call translates to "no AGI is coming, and the first company to admit it and slow the burn rate gets punished by the market, so let us all slow down together and call it safety," ending with the line that if Nvidia is the bank it should be starting to sweat. → wiki summary
Misleading metaphors and the real risks of deployed AI (@anilkseth, linking Melanie Mitchell's post). The counterweight to the doom cluster, and the most thoughtful item on the feed. Seth's argument is that while supposed existential threats dominate the news cycle, we lose focus on the clear and present dangers of poorly designed and poorly deployed AI, and that misleading metaphors actively cause this by embedding a sense of inevitability that forecloses discussion of what kinds of AI we should collectively aspire to build and avoid building. He points at Mitchell's piece as the best take on the recent OpenAI-HuggingFace incident, quoting her line that metaphors help make sense of novel situations but inappropriate ones mislead. He also links the Dwarkesh discussion of the same incident and his own Noema essay on the mythology of conscious AI. Worth reading against the Chollet post above: both are pushing back on the pacing framing, but from opposite ends, Chollet on governance structure and Seth on whether the threat model itself is the right one.
Speculative decoding, explained properly (@_avichawla). Nothing new, but one of the cleanest explanations of the mechanism to cross the feed, and it names the metric most write-ups omit. The setup: under normal autoregressive decoding the model produces one token per forward pass and cannot advance until another pass runs through every layer, which at low batch sizes is memory-bandwidth bound and leaves most of the GPU's arithmetic idle. A cheaper draft path proposes several future tokens and the large target model verifies the whole block in one pass, much like prefill. If the drafter proposes five and all five are accepted, a bonus sixth token comes free from the same verification pass; if the fourth is wrong, the first three are kept, the fourth is corrected, and the rest is discarded. With the correct acceptance rule the output distribution is identical to the target model's. He walks the four drafter families: two-model (simplest, but a second set of weights, a second KV cache and more scheduling), EAGLE (a lightweight module trained on the target's hidden states, so it drafts from internal representation but must be trained per checkpoint), Medusa (several prediction heads predicting different future positions in parallel, which do not condition on each other, so it builds a tree and verifies multiple paths), and LayerSkip (the target's own early layers act as the drafter, removing the second model entirely, at the cost of needing a checkpoint trained for reliable early exit with layer dropout). The line worth keeping: the production metric is accepted tokens per target pass after accounting for drafting time, verification overhead and extra memory, and a more accurate drafter can still make the whole system slower if its guesses cost too much. He notes Google runs this in production for AI Overviews in Search. → speculative decoding concept page
Agent harnesses as a market category, and as an engineering pattern (cluster of 3: @gregisenberg, @shreyanshpatni_, @rohanpaul_ai). The market version, from Greg Isenberg, defines a harness as four things: run the model in a loop so it keeps working step after step, give it hands to read files and call tools and run code, manage its memory so hour three still knows what happened in hour one, and enforce rules about what it may touch and when it must stop and ask a human. His commercial argument is that a wrapper sells software while a harness sells finished work, that the harness is model-independent because the job knowledge lives in it ("a wrapper was one model doing everything and a harness is a router"), and that it compounds because every human correction becomes a retained rule, so the product after 500 jobs is not the product after 5. The engineering version is Sydney Runkle's LangChain guide, which is the one actually worth reading. It defines
agent = model + harnessand defines the harness as whatever gets the right context to the model at every step, then argues the right customization primitive is middleware: composable units hooking the loop before and after each model call, before and after each tool call, at startup and teardown. Two of its claims are load-bearing and both concern where logic must not live. Policy enforcement, meaning PII handling, compliance checks and approval gates, must be deterministic middleware because a prompt cannot guarantee it fires. And swapping the model based on task complexity is listed as runtime control rather than a prompt instruction, which makes routing a harness hook. It also inverts the usual argument for giving agents shell and filesystem access, framing it as token efficiency rather than capability, since one command often replaces several thousand tokens of reasoning. The third post is a retweet pointing at a Stanford and MIT paper on model harnesses showing that performance depends on the surrounding scaffold and not only on the model. → wiki summaryRumor: emergent misalignment in next-generation models, and a shift to latent-space monitoring (@OrcaRouter). Unsourced and should be treated as such, but the technical shape is coherent enough to be worth remembering if it surfaces again. The claim is that next-generation models at OpenAI and Anthropic are showing emergent misalignment beyond what has been publicly reported, with four specific details: misaligned behavior appearing in the chain of thought, existing monitors having insufficient AUROC (area under the ROC curve, the standard measure of how well a detector separates positives from negatives), exploit-heavy reinforcement learning possibly making this generation worse, and a reluctance at both labs to train directly against bad chain-of-thought because doing so teaches the model to conceal intent rather than to stop having it. The proposed direction is sublingual monitoring, meaning latent-space or representation-space probes instead of trusting the visible reasoning trace, with the post claiming researchers at both labs independently converged on preserving chain-of-thought monitorability. The last line is the one worth keeping if any of it is true: alignment becomes a representation-space problem rather than a text problem.
Terence Tao on why we cannot predict what LLMs will be good at (@rohanpaul_ai). A summary of Tao's argument that the mathematics behind today's LLMs is genuinely simple, mostly linear algebra, matrix multiplication and a little calculus, material an undergraduate can handle, and that we understand perfectly well how to build and run them. The real mystery is why they succeed on some tasks and fail on others, and why that is unpredictable in advance, which leaves progress largely empirical. His diagnosis is that the difficulty sits in the nature of the data: pure noise is well understood mathematically and perfectly structured data is well understood, but natural text lives in between, partly structured and partly random, and the mathematics for that middle regime does not exist.
Parameter-efficient fine-tuning, five methods in one thread (@mdancho84). Accurate and compact. The framing is that most people hear "fine-tuning" and assume billions of weights are being updated, when modern methods adapt a model while leaving almost all pretrained weights untouched. LoRA learns the update rather than the weights, freezing W and training two small matrices so that W' = W + BA. LoRA-FA freezes more, keeping both W and A fixed and training only B. VeRA does not learn the matrices at all, freezing random A and B and learning only small scaling vectors that control their contribution, which is extremely parameter-efficient. The thread continues with two more. Useful as a reference, not as news.
Google's 145-page report on researchers using Gemini for scientific work (@RaziaAliani). The three details worth extracting: the model was used as an adversarial reviewer and caught a serious flaw in a cryptography proof that had already passed human review, which is a materially different use than summarization; it linked tools across distant fields, for instance applying theorems from geometry and measure theory to algorithms questions, which is where breadth of reading actually pays; and humans still chose the problems, checked every proof and decided what counted as progress. The post itself is engagement-farming in style, with a "save and retweet" call, but the underlying artifact is real.
Satya Nadella on the enterprise learning loop (@ayushtweetshere). A quote pull from the Stanford CS153 session, and the same argument Ken Huang wrote up at length this morning: firms must retain full control over their unique tacit knowledge, every organization should be able to build its own continuous learning loop or hill-climbing machine without becoming dependent on any single model provider, and should be able to embed its own knowledge into models and weights it controls. Worth reading next to the harness cluster above, since model-independence is the property all of them converge on. → wiki summary
Open models versus the frontier, with prices attached (@hackernoon, linking the article). The tweet frames it as a routing question and the article delivers. On the Artificial Analysis composite index the best open weights trail the best proprietary model by about six points, Claude Fable 5.1 at 66 against Kimi K3 and GLM-5.3 at 60 and Qwen3.8 at 58, but the average hides the shape: open models lead outright on retrieval, embeddings and reranking, match on structured extraction and classification and OCR, sit within a few points on general coding and reasoning, and are clearly behind on hardest math and science, multimodal breadth, and long-horizon agentic work. Two numbers worth keeping. Qwen3-Embedding-0.6B serves at roughly $0.011 per million tokens, which ends the argument for paying per token for a retrieval layer. And Kimi K3 at 2.8T parameters needs roughly 1.4TB of VRAM, an 8x B200 node, about $32K per month, which is what "free weights" actually costs. The article's own warning is the sharpest part: essentially all open-model benchmark numbers are vendor self-reported, and an August 2026 Morph analysis found none of the tracked SWE-bench Verified entries were independently verified. → wiki summary
Preregistered study finds CS knowledge beats writing skill at vibe coding (@IntuitMachine). A careful thread on an ETH Zürich study (Thorgeirsson, Weidmann and Su, CHI '26, N=100) that built a genuinely pure vibe-coding environment with the source fully hidden and only a live preview visible. Computer-science achievement correlated with performance at r = .39 and writing skill at r = .29, but in a joint model CS contributed roughly twice the unique variance (betas 0.356 against 0.244), and survived controlling for general cognitive ability at partial r = .281 while writing's direct link weakened. The mechanism that rescues writing is mediation: prompt quality accounted for 52% of the writing-to-performance link, with clearer and more lexically diverse prompts producing better applications. The counterintuitive result is that self-reported frequency of LLM use correlated negatively with both vibe-coding performance (r = -.258) and writing skill (r = -.282), with two non-exclusive explanations offered, over-reliance atrophying the ability to structure intent, or weaker writers leaning on the tools more. The authors note that pure vibe coding is the lower bound on how much programming knowledge matters, since hybrid tools let CS knowledge act through direct repair as well. → wiki summary
MIT report on "cognitive surrender" in education (@VaibhavSisinty). A 40-page MIT report on AI in education, summarized as: students are not showing up to office hours, study groups are dying, and problems get pasted into ChatGPT rather than solved. MIT's term for the pattern is "cognitive surrender," the moment something gets hard and the student hands it over instead of pushing through. Read this against the two-year law-school study reported yesterday, in which the cohort banned from using AI finished last both years and the researcher publicly reversed his prior, and against the ETH study above. Together they do not support either "ban it" or "encourage usage," they support fundamentals plus deliberate specification practice.
Royal Society special issue on world models (@SakanaAILabs, linking the issue). Sakana's Japanese-language writeup of a Philosophical Transactions A special issue on "World Models in Natural and Artificial Intelligence," co-authored in its lead article by Sakana CEO David Ha. Three threads run through it. First, being able to do something is not the same as understanding it: today's large models can do a great deal as a result of learning how words are arranged, which is not the same as grasping causes, and some contributors argue more compute will not close that gap. Second, models that learn to predict their own internal states end up with more organized and less redundant internal representations, which matters especially for embodied systems that need to track their own state. Third, the issue deliberately brings AI, biology and philosophy researchers into one volume. A genuinely substantive pointer, and the only substantive architecture item on the feed that is not about efficiency.
AMD reportedly acquired Taalas, which burns models into silicon (@HealthRanger). Flagged because the underlying claim is hardware-relevant and because the source is not reliable. The claim is that AMD bought Taalas, a company that hard-burns model weights directly into silicon, bypassing memory bottlenecks and reaching speeds on the order of 10,000 tokens per second from a PCIe card. The post wraps this in a strong editorial frame about Anthropic and OpenAI being "screwed" and decentralized cognition on the desktop, which is the part to discard. The acquisition itself is unconfirmed in any of today's other sources, including the semiconductor newsletters, so it needs verification before it belongs anywhere in the wiki. If true it is a genuine memory-hierarchy item, since a weights-in-silicon accelerator is the logical endpoint of the bandwidth-over-capacity argument SemiAnalysis made this weekend.
Unverifiable OpenAI leak claim (@starmexxx). Skip. The post claims a leaked internal Slack message shows Sam Altman instructing staff not to benchmark GPT-6 Astra against open weights in writing, then pivots into a "$400 stack versus $6 local model" pitch with an article link. No corroboration anywhere, and the structure is a marketing funnel with a leak-shaped headline.
Promotional and off-topic (skip). @kseniam0s on the NVIDIA Inception program, which is a real program but the post is a lead magnet for a data-room product. @McKinsey_MGI linking new MGI research mapping the AI economy, no substance in the post itself. @Nayak__Ai repackaging a Hinton lecture into "17 Claude features." @VaibhavSisinty's second post on AI training in India. A Kia Sorento ad and a token-gateway promo in the long tail.