Summary
The afternoon slot carried 44 tweets, no @bayesiansapien retweets for a sixth consecutive day, and about three quarters of the volume was off-topic political and culture content. One item is worth real attention: DHH ran the same well-specified task, a full Rust rewrite of the TerminalTextEffects Python animation library, through three different frontier models and published the token and dollar cost of each. Claude Fable 5 one-shotted it in 11M tokens and produced a 9.6x rendering speedup with startup time down from 87ms to 2ms. Grok 4.6, handed Fable's plan rather than the raw task, finished in 1h24m on 8.6M tokens for roughly $55, which DHH puts at about a tenth of the Fable cost. GPT Luna ran over eight hours, got one of 37 effects working, and he stopped it. That is a rare public three-model cost-per-completed-task datapoint on a single non-toy job, and it is the closest thing this slot has to a measurement. Pairing with it, Meta AI's Nikita Skorobogatov argues that spawning a cloud agent per task is straightforwardly wasteful for small edits and that git worktrees degrade with repo size, so the right fix is kernel-level rather than more remote compute. The rest is thin: one post on Colossus compute versus algorithmic talent as the real bottleneck, a complaint that AI-generated replies on X add no value, and a large block of geopolitics from @MarioNawfal that carried nothing AI-relevant.
Posts
- One task, three models, three very different bills (cluster of 2: @dhh, @dhh · artifact). DHH rewrote the TerminalTextEffects Python library in Rust with agents and reported the economics of each attempt: Fable 5 one-shotted it in 11M tokens for a 9.6x rendering speedup, 87ms to 2ms startup, zero dependencies and a 3mb single executable; Grok 4.6, given Fable's plan instead of the bare task, reproduced the result in 1h24m on 8.6M tokens at roughly $55, about a tenth of the cost; GPT Luna ran more than eight hours, landed one of 37 effects, and was stopped. The cheap-model-plus-strong-plan result is the interesting one, because it separates the value of the planning pass from the value of the execution model, which is exactly the decomposition a router would need. Relates to agent harness engineering and LLM routing.
- Spinning a cloud agent per task is compute you are throwing away (@theskory). Skorobogatov's point is that using a remote agent for something as small as renaming a variable wastes cost and compute on cloning, snapshot restore and cold start, and that git worktrees were never designed for isolation at scale so they get worse in proportion to repo size. His prescription is to squeeze the kernel and OS for cheap isolation rather than reach for another cloud instance, which is a cost-side counterweight to the current default of provisioning a fresh environment per agent. Relates to compute economics.
- Agents make better bug reports than users do (@dhh). Omarchy Quattro routes bug reporting through agents, and DHH says the result is fewer reports of much higher quality. Small observation, real mechanism: the agent can reproduce, bisect and attach context before filing, which shifts triage cost from maintainer to machine.
- Compute is not the bottleneck, the training recipe is (@ns123abc). The argument is that SpaceXAI already has the Colossus compute and the Cursor coding-data recipe, so its remaining gap to the frontier is algorithmic: scaling to 10T parameters needs pre-training and RL breakthroughs that only a very small number of people can produce, citing the three-person Olympiad team behind OpenAI's 2025 IMO gold model. Unsourced and self-serving in places, but the underlying claim, that marginal frontier progress is now talent-bound rather than FLOP-bound, is a testable one.
- AI-generated replies on X are vapid enough to be an industry embarrassment (@Scobleizer). Scoble's complaint is that the reply bots restate what you said and add nothing, and his framing is that genuine capability would show up as good comments rather than fluent ones. Worth noting only as a reminder that the visible surface of deployed AI to a general audience is still the low end of it.
- Matic robot vacuum ships voice control, two years after the demo (@Scobleizer). You can now tell the robot where to clean. Consumer robotics product update, one line of signal.
- Taste becomes the scarce skill when intelligence is abundant (@brivael). Knowing what to make, what to pick, and what to say is the residual differentiator once generation is cheap. A slogan rather than an argument, but it is the same thesis DHH's experiment demonstrates in practice, where the plan carried more value than the executing model.
- X creator payouts, disclosed (cluster of 3: @brivael, @brivael, @brivael). Over $1,500 a month from subscriptions on a large account, with an audience split of roughly 30% US, 30% France, rest of world, and a new program expected in September. Platform economics, not AI.
- Omarchy Quattro RC3 running on a PlayStation 4 Slim (@dhh). Hobbyist port, no AI content.
- French politics, media and assorted reposts (cluster of 9: @brivael, @brivael, @brivael, @brivael, @brivael, @brivael, @brivael, @brivael, @dhh). Macron's term limits, state spending cuts, MeToo commentary, Manhattan rents at a record $6,655 a month, a piano stream, and a "best part is no part" repost. Nothing AI-relevant.
- Geopolitics and culture block (cluster of 20: @MarioNawfal and 19 further posts in the same run). Hormuz and the threatened Iran naval blockade, Brent at about $88, drones intercepted over Erbil, US immigration enforcement, UK asylum housing, an Alabama execution, plus assorted viral clips. High volume, zero AI signal, and it dominates this slot's raw count. Skip.
- Assorted one-liners (cluster of 3: @JonasBadalic, @spencerpratt, @zhu_hanqing666). "Deploy straight to prod or bust", a Netflix trailer reaction, and an emoji reply. Skip.