Summary
The morning slot carried 24 tweets, no @bayesiansapien retweets for a fifth consecutive day, and roughly a third of the volume was off-topic political content. The strongest signal is a two-post cluster on the DeepSeek Harness release: Hugging Face's Elie Bakouch, reading the code, reports that its central engineering commitment is KV-cache-aware design, meaning previously written conversation history is never edited and modifications are appended at the tail instead, because editing a prefix invalidates every cached token after it. He also estimates at least 20% of the harness's own commits came from Codex worktrees, so the harness was substantially built by agents. The second cluster is Cognition, with six posts spanning two unrelated things: Gemini 3.7 Flash landing in Devin at Claude Sonnet 5 parity on their internal FrontierCode 1.1 benchmark for under half the cost, and xAI publishing the actual weights of the X recommendation algorithm, where the single largest positive signal turns out to be a share via copy link at 20.0 and the largest negative is a report at -234.0. NVIDIA ran six posts, the substantive one being LG building a next-generation bipedal humanoid on Isaac GR00T alongside a chairman-level visit expanding their AI infrastructure and robotics work, plus a "tokens are the new commodity" tokenomics push. Two smaller items worth noting: Matt Shumer has had a self-running Gauntlet Loop on Grok 4.6 going for over a day and observes that not all models can sustain that, which is a practitioner data point on harness-bound versus model-bound behaviour, and Devin can now query and provision MongoDB Atlas databases mid-task, removing the stale-schema problem from data-agent work.
Posts
DeepSeek Harness is KV-cache-aware by construction, and was largely written by agents (@eliebakouch, repo). Bakouch's teardown is the most technically substantive post in the slot. DeepSeek Harness is a web UI containing multiple harnesses, able to spawn Claude Code and Codex agents through their SDKs, and it ships several default modes which are really different harnesses: code mode with programmatic tool calling in TypeScript, bash-and-edit (the mode usually used in evals), and a standard read/write-tool mode. The design detail that matters: there is first-class KV-cache-aware design, meaning the harness makes sure the KV cache (the store of previously computed attention keys and values that lets a model skip recomputing tokens it has already seen) is never altered. For every change, it does not modify the previous history, which is the prefix; it appends something at the end stating the modification. Bakouch expects other harnesses to follow. He separately estimates at least ~20% of the harness's commits and PRs came from Codex worktrees, adding that he counted only worktree or named-branch signals so the real figure is probably higher. The cost significance is covered in today's digest and the DeepSeek cache-economics summary: DeepSeek repriced cache-hit tokens roughly six-fold the same day, so the harness is the mitigation for the pricing.
Gemini 3.7 Flash reaches Claude Sonnet 5 performance in Devin at less than half the cost (cluster of 2: @cognition, @cognition, blog). Cognition put Gemini 3.7 Flash into Devin Desktop and Devin CLI and published its position on FrontierCode 1.1, their proprietary benchmark that grades models on real engineering tasks by quality and mergeability, with solutions that fail blocking criteria scored zero. The numbers on that board: Gemini 3.7 Flash 56.3, Claude Sonnet 5 56.2, GPT-5.6 Terra 56.2, GPT-5.5 56.7, Kimi K3 58.2, GPT-5.6 Sol 60.6, Grok 4.6 61.3, Claude Opus 5 63.6, Claude Fable 5 64.9, with Claude Sonnet 4.6 trailing at 40.0. So Flash ties Sonnet 5 at roughly half the price rather than beating the frontier. The operationally useful detail is narrower and more specific than the headline: inside Devin it performs particularly well on tightly scoped refactors, where it delivers minimal diffs that match repo conventions. That names a task class where the cheap model is preferred rather than merely tolerated, which is a routing signal rather than a marketing one. Available at an extra 50% discount through 2026-08-27. Relates to LLM routing.
xAI released the X algorithm weights, and the largest positive signal is sharing a link (cluster of 3: @cognition, @cognition, @cognition, DeepWiki index). The attached screenshot is a DeepWiki table of ranked signals with default weights, drawn from
home-mixer/scorerin thexai-org/x-algorithmrepo. Reading it top to bottom: share via copy link (ShareViaCopyLinkWeight) 20.0, reply (ReplyWeight) 5.0 with a +15.0 boost to 20.0 for bidirectional-follow original posts, share via DM 5.0, quote 5.0, follow author 4.0, share 2.0, retweet 1.0, favourite 0.5, click 0.4. The negative side, posted separately: not interested -43.2, mute author -58.8, report post -234.0. Two things are notable for anyone who thinks about ranking systems. The positive-signal scale spans roughly 50x from a click to a copy-link share, so the system is overwhelmingly optimizing for off-platform propagation rather than on-platform engagement. And the negative weights are an order of magnitude larger than any positive one, so a single report costs more than eleven copy-link shares earn, which is a very aggressive suppression posture. Cognition's own contribution is the DeepWiki index of the repo, which makes the codebase browsable rather than requiring a clone.Devin now works directly against live MongoDB Atlas data (@cognition). The agent can query a live database, manage it, and provision a new cluster mid-task, with the pitch being no stale schemas, no copied-in context, and no waiting on a human to set up a database. Worth noting because copied-in schema context is one of the standard failure modes in data-agent work: the agent reasons correctly over a snapshot that no longer matches production. This moves the schema from context to tool call, which is the same "take the decision away from the model, put it behind a deterministic interface" pattern the harness research keeps landing on.
A Gauntlet Loop has been running unattended on Grok 4.6 for over a day (@mattshumer_). Shumer reports a self-running loop going for over 24 hours and calls it super promising, with the observation that not all models can do this. That last clause is the practitioner version of the model-bound versus harness-bound diagnostic: the loop is the same, and whether it survives a day depends on the model's ability to avoid drifting or wedging. No numbers, no output yet, so it is an anecdote rather than evidence, but it is the kind of anecdote that precedes a measurement.
LG is building a next-generation bipedal humanoid on NVIDIA Isaac GR00T (cluster of 2: @nvidia, @nvidia, release). LG announced development of a next-generation bipedal humanoid built on Isaac GR00T, NVIDIA's open reasoning humanoid stack, and NVIDIA separately hosted LG Group chairman Kwang-mo Koo and LG leadership to mark an expansion of their collaboration across AI infrastructure, physical AI, and robotics. The robotics substance sits outside this wiki's attention range, but the partnership shape is worth logging: another large consumer-electronics manufacturer standardizing on NVIDIA's robotics platform rather than building its own, which is the same platform-lock dynamic that makes the CUDA fungibility argument on the compute economics page work.
NVIDIA is pushing an AI-factory and tokenomics framing (cluster of 2: @nvidia, guide). The pitch is that AI factories are the industrial infrastructure of the AI era, tokens are the new commodity, and the question for an operator is how to optimize token economics so compute becomes revenue. Treat it as a vendor artifact rather than an analysis, but the framing is the same one the day's research is converging on from the other side: if tokens are the priced commodity, then every result that reduces tokens per completed task is a revenue-side lever and not just an engineering nicety. Worth reading against the AlphaSense finding that token price and task cost can point in opposite directions, which is precisely the complication a tokenomics guide has an incentive to leave out.
Runway brought Gen-4.5 up on NVIDIA's Vera Rubin platform in one day (@nvidia, Runway). At the 2026 Runway AI Summit, NVIDIA's Richard Kerris argued AI is a new compute model for media and entertainment, with the chain being that real-time rendering made simulation possible, simulation made generative AI possible, and agents you talk to turn the story itself into the interface. The concrete number is the porting time: Runway had Gen-4.5 running on a brand-new hardware platform in one day, against eight years from RTX to real-time rendering. That one-day figure is the CUDA-continuity argument as a data point rather than a claim, and it belongs with Jensen Huang's fleet-fungibility thread from yesterday afternoon.
Speculation that xAI is circling Cognition for the talent (@ns123abc). The argument: Elon has Colossus compute and the user coding data recipe from Cursor, and now needs the most technically capable team to write the next training recipe, which the poster claims is unusually concentrated at Cognition (Scott Wu, three-time IOI gold, and colleagues), citing OpenAI's 2025 IMO gold model coming out of a three-person olympiad team as precedent. Unsourced speculation with no confirmation anywhere in today's raw sources. Logged because the underlying claim, that a small dense team of competition-mathematics people produced a capability breakthrough that shifted the whole field's approach to reasoning models, is a recurring narrative worth watching for evidence rather than repetition.
Grok 4.6 usage limits reset during launch, plus a sketch-to-website demo (cluster of 2: @theskory, @minchoi). xAI reset usage limits via a token in desktop and mobile settings to keep people building through the Grok 4.6 launch, and a widely shared demo shows an input sketch turned into a finished website from the single instruction "change this into website" with no skill file and no special instructions. Product signal, not research signal, but the no-skill-no-instruction framing is the interesting part: it is a claim that capability has moved from the harness back into the model for this particular task, which is the opposite direction from everything else in today's digest.
Off-topic and promotional, skipped: @DoWCTO on drone-production tariffs, @MarioNawfal on a Joe Rogan podcast clip, @spencerpratt on California electricity policy and HUD funding (two posts), @Scobleizer on being reshared by Elon, @dhh on open-source social life, and @eliebakouch's bare cc mention. Roughly a third of the slot's volume, none of it AI-relevant despite passing the keyword filter.