Summary
The strongest signal in today's slot is not a debate, it is a benchmark table nobody asked for: tinygrad published a complete serving profile for DeepSeek V4-Flash-0731 on two RTX Pro 6000 Blackwell GPUs, hitting 245 tokens per second sustained for a single user and beating the model's own runbook figure, then immediately launched a 2-GPU tinybox variant off the back of it. Second in weight is @ns123abc's breakdown of OpenAI's Astra announcement, which is the only place in the slot where the actual list of ten solved mathematical problems appears with the Lean proof repository attached. Third is Mira Murati amplifying Thinking Machines' "A Safe Path to Open Weights," the first written-down open-weights release procedure any lab has published, whose full text the farmer captured. There is a small but real cluster of three posts (cluster of 3) on agents becoming hard for humans to supervise: an xAI engineer arguing agents overcompensate with sophisticated-sounding answers, DHH shipping a Linux desktop panel that tracks Claude and Codex subscription burn, and Min Choi publishing a prompt that tries to make an agent stop asking questions it could answer by reading the repo. NVIDIA's one substantive post is a customer case study claiming roughly 10x lower inference cost. The curated retweet feed from @bayesiansapien was empty in this slot, so the AI handle feed carried everything, and roughly two-thirds of it was geopolitics, US domestic politics and French commentary from accounts that are only nominally AI-adjacent. Signal density was low; the four posts that mattered mattered a lot.
Posts
tinygrad publishes a full DeepSeek V4-Flash serving profile and gets 245 tok/s on two GPUs (@tinygrad). The attached image is a benchmark table, and it is the most useful artifact in today's slot. Running DeepSeek-V4-Flash-0731 with tensor parallelism across only 2 of the box's 4 GPUs (physical GPU 0 and 1, about 95 GB used each, GPUs 2 and 3 idle), the numbers are: single user, 128 in and 512 out, about 245 tok/s sustained, with median time-per-output-token around 2.8 ms translating to 300 to 350 tok/s instantaneous, and time-to-first-token about 0.4 s. Sixteen concurrent streams: 608 tok/s total at 17 ms per stream, with speculative acceptance at 63.8%. Sixteen concurrent streams at long context (8192 in, 512 out): 499 tok/s decode, 8,477 tok/s total including prefill, with acceptance at 90.5% and average accepted length 5.52. The configuration line is the part worth copying: vLLM "Gilded Gnosis" r16, B12X W4A8 kernels (four-bit weights, eight-bit activations), DSpark K5 fixed-depth speculative decode, fp8 KV cache (the attention memory store held in eight-bit floating point rather than sixteen), 131k max context, temperature 0. Two honest footnotes carry more information than the headline. First, 245 tok/s single-user beats the runbook's own validated 217 to 220 tok/s, so this is a practitioner outperforming the vendor's published figure. Second, and more transferable, the 90.5% long-context acceptance is inflated by repetitive random-token prompts and real code is approximately 64%, a 27-point workload gap on the same model and the same scheme, which quietly invalidates any speculative-decoding speedup quoted without its prompt distribution. See today's Deep Dive.
tiny corp turns the benchmark into a product the same hour (@tinygrad, product page). Because the DeepSeek run only needed two GPUs, tinybox green v2 now ships in 2-GPU and no-GPU variants alongside the 4x RTX Pro 6000 Blackwell configuration at $31,000 (384 GB VRAM, 32-core AMD Genoa, 192 GB system RAM, 4 TB raid array, two 1600W PSUs, 90 lbs). All variants include the full hardware to install four GPUs later, and both include two power supplies, so the upgrade path is drop-in. Made to order, 2 to 8 weeks to ship. The pricing logic is a direct consequence of the benchmark: if a 304B-parameter model serves at interactive latency on two cards, the four-card box is over-specified for that workload and tiny corp would rather sell the two-card box than lose the sale.
OpenAI's Astra announcement, with the actual list of ten results (@ns123abc, source tweet, OpenAI post, Lean proofs). The enumerated claims: non-sofic groups constructed, Connes's rigidity conjecture disproven, sphere packing with the Cohn-Elkies bound strength determined exactly, quantum parallel repetition proven for all entangled games, closest-vector-problem hardness, Ehrhart's volume conjecture, permanent lower bounds, and three Erdős problems resolved (#146, #180, #183). The line that makes this different from every previous AI-mathematics claim: four of the ten are counterexamples to what mathematicians believed, and every one ships with a Lean 4 formalisation in a public repository, so a referee runs a type-checker rather than trusting a natural-language argument. The repository contains one
.leanfile per result (NonSoficGroup.lean,ConnesRigidity.lean,SpherePacking.lean,GapCVP.leanand so on) but its README does not state whether the formalisations are complete, how the proofs were produced, or what was human-verified, which is the first thing anyone should check. Astra itself is a model family for long-running multi-agent work over hours or days, tokens for all ten solutions cost roughly $2,000 at API rates, and it will be the first model routed through a new US federal pre-release review. See today's Deep Dive.Mira Murati amplifies the first written-down open-weights release procedure (@miramurati, Thinking Machines post). The framing in the quoted Thinking Machines post is the sharpest sentence any lab has produced on this: "Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a path between them. We haven't mapped all of it." The substance behind it, from the full post the farmer captured: two gating questions (is the model safe, is the ecosystem ready), internal evaluations across chemical/biological/radiological/nuclear risk and cybersecurity plus multimodal content in 17 languages, external red-teaming by four organisations with deliberately disjoint mandates (Scale AI on general misuse, Handshake AI on vulnerable populations, FAR.AI on CBRN and cyber, Apollo Research on scheming and sabotage), and adversarial fine-tuning that strips safety training to measure the model that will exist a week after release rather than the one shipped. None found capability exceeding existing open-weight models. The proposal is staged access rather than a binary: inference API for defenders, then fine-tuning for vetted defenders through Tinker, then white-box access for safety researchers, then monitored public access, then full release, with progression conditional on evidence. The load-bearing technical claim is that dangerous capability may be decouplable from general intelligence, because much dangerous knowledge is looked up rather than derived, so filtering CBRN documents from pretraining drops harmful-capability scores while leaving unrelated capability intact. See today's Deep Dive.
Agents are outrunning the humans supervising them (cluster of 3) (@JonasBadalic, @dhh, @minchoi). Jonas Badalic of xAI responds to Santiago's complaint that agents are getting harder to understand ("my mind is no longer able to keep up with the massive context they work with and the summary output feels like reading another language") with a diagnosis and a test: agents "try really hard to solve your task and overcompensate by giving you a sophisticated sounding answer, same as humans," and you can prove it by asking for an explain-like-I'm-five, where you will often get two sentences that are completely wrong. DHH is attacking the same problem from the instrumentation side, shipping an Omarchy Quattro panel that auto-detects Claude and Codex subscriptions and surfaces usage against limits. The screenshot shows the Codex Pro tab with weekly limit consumption at 7%, a reset countdown of 4 days 17 hours, a per-day token bar chart (Wednesday at 109.7M, today at 23.8M, Sunday 3.3M, Monday and Thursday zero) and a per-model breakdown showing GPT-5.6 Sol at 165.5M tokens. That is a burn-rate dashboard for an agent, and the fact that a desktop environment now ships one as a first-class panel is a better indicator of where agentic coding actually is than any benchmark. Min Choi publishes the third response, a prompt preamble instructing the agent to "work like a contractor who bills for rework," investigate before asking, and never ask about test framework, language version, lint rules, error-handling conventions, directory layout or existing abstractions that are discoverable in under a minute of searching. Three different people, three layers of the same problem: the agent's output is unreadable, its cost is invisible, and its questions are wasteful.
NVIDIA claims a customer cut its largest inference workload roughly 10x (@nvidia, case study). StudyFetch transcribes hundreds of thousands of college lectures a month and reports roughly 10x lower cost using NVIDIA Riva, Parakeet ASR and NIM microservices, with the savings funding voice tutoring, real-time personalisation and a new agentic learning platform. Vendor-published and unaudited, and the mechanism is unstated beyond the product names, so read it as a direction rather than a number. It is worth logging because speech-to-text at that volume is one of the few inference workloads where the unit economics are public enough to check later.
Grok Imagine 1.5 adds text-to-video and omni-reference (@hexiang, quoting @grok). xAI's video model gains text-to-video support, image and voice references, and native 1080p. Xiang He of Google DeepMind, notably a competitor, is the one amplifying it, and his read is specific: "with omni-reference, now imagine 1.5 is much easier to use for making long story with consistent characters." Character consistency across shots is the binding constraint on video models being usable for anything longer than a clip, so a reference-conditioning mechanism is a more meaningful upgrade than a resolution bump.
A local Gemma 4 on a phone coached a race car in real time (@Scobleizer, quoting @antigravity). At Sonoma Raceway, Google Developer Experts built an AI race coach bridging vehicle telemetry and application code, tapping over 100 car sensors through a Pixel 10 with Gemma 4 running locally on the device to deliver real-time audio coaching. Robert Scoble's addition is the interesting claim rather than the demo: he says he watched AI beat a human driver on a track in Abu Dhabi. The technically notable part is the deployment shape, which is a small open-weight model on a phone in a hard-latency loop with sensor input, not a cloud call.
Xiaomi's factory humanoid reports 98% accuracy on a real station (@Scobleizer, quoting @spaceandtech_). After four months at a self-piercing nut station in Xiaomi's Beijing EV factory, the robot reached 98% task accuracy, and is now also handling centre-console side-cover sorting and returnable-box folding at over 90% each. Scoble's one-line comment is "not available in USA." Four months of continuous deployment on one station with a hard accuracy number is a substantially more credible robotics claim than a demo video, and the progression to two additional task types is the part that would matter if it holds.
The US Department of War commits $820 million to domestic drone component manufacturing (@DoWCTO, press release). The Office of Strategic Capital signed a conditional loan commitment of up to $820 million with Performance Drone Works to establish high-volume domestic manufacturing capacity for Group 1 and Group 2 unmanned aircraft components. Framed explicitly as addressing supply-chain bottlenecks. It is one of the largest drone financing deals in US history and belongs here as compute-and-hardware-adjacent industrial policy rather than as AI research.
Leopold Aschenbrenner writes to his limited partners (@ns123abc, quoting @tbpn). TBPN published the full letter Aschenbrenner sent his LPs after Situational Awareness unloaded nearly its entire public portfolio to Citadel following margin calls, days after reporting a 439% six-month return. The letter's line is that rumours of the fund's demise are exaggerated. Nothing new on the mechanics beyond what was logged yesterday, but the letter itself is now public.
"Agents are getting so good" (@stepango). A four-word post from an xAI engineer with no elaboration. Logged for completeness because it is the third xAI-affiliated post in the slot on agent capability, all of them impressionistic rather than measured, which is itself a small signal about how the inside view is being expressed right now.
Skip. Roughly two-thirds of today's AI handle feed was off-topic for this wiki: @MarioNawfal (16 posts on Ukraine, Iran, Ceuta, Red Sea shipping and Australian wildlife), @brivael (19 posts of French political commentary on migration, natalism and EU institutions), @NICKIMINAJ (8 posts on Fauci), @spencerpratt (7 on California politics), @heavypulp (6 on song lyrics and a memecoin), plus @AustinJustice, @SeanParnellASW and @lynnmartin. Also skipped: @Scobleizer's XR glasses and VITURE posts, @dhh's wifi QR code panel, and @stepango's Raptor engine test-fire repost, none of which carry AI substance.