AI Brief, 25 October 2026: a robot policy for bodies it has never seen

Dream Labs came out of stealth on 19 October with $90 million in seed funding and a claim that matters more than the money: its first model, Dream-1, can control two-armed robots it was never trained on, with no demonstrations from those robots at all. The launch went uncovered here for six days and reached the Hacker News front page only on the 24th, at 157 points. It leads late rather than not at all.

What makes it reportable rather than another robotics demo reel is that somebody else measured it. Robocurve, a public-benefit corporation publishing frontier robotics evaluations, scored Dream-1 at 59.3% on RoboDojo-RC Tier 1 over 120 attempts per configuration, against 48.2% for the strongest large-model configuration on its board, Gemini 4 Argon, and 38.2% for Claude Opus 5.5, at $0.15 per attempt against Gemini 4 Argon's $2.34. Robocurve publishes that table on its own site, so the figure does not rest on the launch post citing it.

Two other items are worth the reader's time. A Linux kernel module published on 18 October, and found by a wider audience on the 24th, restores the deliberately halved BF16 matrix throughput on consumer Ampere GeForce cards, with user-reported gains of 25% to 81% on dense GEMM and bit-identical outputs. And Apollo's chief economist Torsten Sløk argued on the 24th that AI has delivered "zero financial payoff to corporate America outside of tech", which drew 513 points and 530 comments, the loudest new AI item of the day.

• Dream-1 is a World Action Model behind a public API at $3 per hour, or $2 for a flash variant. Each call returns 1.6 seconds of actions, so an hour is 2,250 calls and about $0.0013 per action.
• Zero-shot tested on 27 unseen robots in simulation and four real ones, including the original PR2, released eighteen years ago.
• Its own demo page discloses attempt counts: one of the showcased tasks succeeded once in nine tries.
• gpu-unlock takes a 3090 Ti from 87.50 to 157.41 BF16 TFLOPS on an 8192³ GEMM. LLM prefill gains 14–36%; decode does not move, because decode is bandwidth-bound.
• Brotli CVE-2026-111497, CVSS 9.2, is a bidirectional unauthenticated RCE that OpenAI says it found in February. The root cause was fixed in source then without being recognised as a security bug.
• Sløk's case: revenue per worker and margins rising for the Mag 7, flat for the S&P 493.

A shared action space, and video as the thing every robot has in common #

Dream Labs was founded this year by former leads of NVIDIA's GR00T robotics research; its chief executive, Joel Jang, led the world-model team on that project. The $90 million seed was announced at 23:06 UTC on 19 October, led by B Capital with Kindred Ventures, and covers more than twenty hardware partners on robots costing roughly $2,000 to $10,000.

The mechanism has two parts, and the first is unglamorous. Robot policies usually emit joint angles, which binds them to one body. Dream-1 can instead emit end-effector poses: where each gripper should be and whether it is open or closed. Every two-arm parallel-gripper robot shares that action space, and each robot's own inverse-kinematics solver turns a pose into joint commands.

The second part is why the transfer works at all. A World Action Model predicts video and actions jointly, so acting and imagining are the same computation, and every robot therefore arrives in the same format: pixels from a head view and a wrist view on each arm. The head camera sees an unfamiliar gripper as shapes moving through a scene. Earlier work on the Universal Manipulation Interface achieved zero-shot transfer across arms, but only when gripper hardware matched exactly and only from wrist cameras, discarding the scene view. Predicting video instead of matching hardware removes that constraint.

Getting started is three API calls, which is the product rather than a detail of it:

import dream

client = dream.DreamClient(api_key=API_KEY)
client.policy({"checkpoint": "dreamlabs/dream-1"})
client.set_observation(observation)   # three camera frames + a plain-English instruction
action = client.get_action()           # 1.6 s of end-effector poses, ~$0.0013 a call

Simulated evaluation ran 300 rollouts per embodiment across 27 robots, all reported relative to YAM, the robot Dream-1 actually trained on, which is the right baseline and also a ceiling. Robocurve's board carries one methodology caveat: it caps large models at 40 model calls per attempt and vision-language-action models at 3,600 control steps, and the VLAs always run to their cap while a large model can stop early.

Dream-1 sits on the Pareto frontier for both cost and time, at 1.5 minutes per attempt against 3.1 for Opus 5.5, and the $2.34 against $0.15 is the 16× the launch post claims. One figure on that post does not check out: its chart caption reads "59 on RoboDojo-RC Tier 1 against 48 for Opus 5.5 and Gemini 4 Argon", but 48.2% is Gemini 4 Argon's alone, and Robocurve puts Opus 5.5 at 38.2% and 36.0% across its two configurations. The error runs against Dream Labs, crediting a rival with ten points it did not score.

Where zero-shot is not enough, Dream Labs sells fine-tuning: five minutes of a robot's own data lifted language-following on ten of twelve beta-partner robots, while a Dexmate with Sharpa dexterous hands took 77 minutes. The company states the limit plainly: "One-shot adaptation still outperforms zero-shot, so a gap remains." Nothing here is open weights, and nothing beyond Robocurve's board is independently reproduced.

The halved rate on consumer Ampere, and why it only helps prefill #

A kernel module by a developer working as Zagitta restores the full BF16 and FP16 matrix issue rate on GeForce Ampere silicon. GA102, GA103, GA104 and GA106 cards issue half-precision FMA into FP32 accumulators at half the rate of their professional equivalents, and that turns out to be a configuration register rather than missing hardware. The repository dates from 18 October, was last pushed at 15:47 UTC on the 24th, and reached 207 points on Hacker News that morning; it stands at 3,433 stars and 60 open issues, mostly compatibility reports.

It writes two speed selectors, SS0 = 0x88888888 for the full BF16 rate and the 4-bit SS1 = 0x8, then performs a function-level reset before the unmodified NVIDIA driver binds the card. Doing that from an early kernel module rather than a UEFI application or a driver patch is the practical advance: it survives driver updates without re-patching.

On a 3090 Ti, dense 8192³ GEMM at 450 W, five-sample medians with FP32 accumulation:

	Before	After
FP16	87.46 TFLOPS	118.39 TFLOPS
BF16	87.50 TFLOPS	157.41 TFLOPS
TF32	42.58 TFLOPS	82.61 TFLOPS

BF16 nearly doubles; FP16 gains 35%. Across user compatibility reports the BF16 GEMM gain is +68–80% on a 3090 Ti, +25–67% on a 3090, +70–73% on a 3080 10 GB, +73–81% on a 3070 and about +80% on a 3060 12 GB, with the power limit explaining most of the spread and the +25% coming from cards capped at 225 W. An OEM 3080 Ti 20 GB completes the unlock and stays at half rate; no GA103 report exists yet. Every figure is a user's own benchmark, so none is independent, though the reports that checked found outputs bit-identical before and after.

The part worth internalising is which workloads move. Prefill gains 14–36%; decode does not move at all, and that is arithmetic rather than a bug. Let $B$ be batch size, $L$ the context length and $s$ the bytes per stored weight. A weight matrix is loaded once and reused across every token in flight, so prefill gets $O(BL)$ multiply-accumulates out of each weight element while decode, processing one token per sequence, gets $O(B)$ . Arithmetic intensity in FLOPs per byte of weight traffic is therefore

$$I_{\text{prefill}}\approx \frac{2BL}{s},I_{\text{decode}}\approx \frac{2B}{s}$$

A 3090 Ti has 1,008 GB/s of memory bandwidth, so at 157 TFLOPS it needs roughly 156 FLOPs per byte to stay compute-bound. BF16 weights make $s=2$ , so decoding at batch 8 supplies 8, a factor of twenty short: the card was already idling on memory and a faster multiplier changes nothing. Prefill at batch 1 and 4k context supplies about 4,100, comfortably compute-bound, and that is where the 14–36% appears. INT8 is untouched. So: worthwhile if you fine-tune, batch, or push long prompts through a pile of 3090s, and no help for single-stream generation.

Sløk puts the payoff outside tech at zero #

On 24 October, Torsten Sløk, Apollo's chief economist, argued that AI "has delivered zero financial payoff to corporate America outside of tech". His case rests on three series: revenue per worker, rising for the Mag 7 and flat for the S&P 493; profit margins, the same split; and aggregate productivity growth, which shows no inflection at all. His conclusion is that mainstream corporate America now treats token spending "not as a source of growth but as a cost center that weighs on margins", with heavy consumption confined to venture-funded companies.

It drew 513 points and 530 comments, more than any other new AI item in the window. It is also a sell-side macro note presented as five charts whose underlying series are not published alongside, so the evidence is an argument from aggregates rather than something anyone can re-run. Flat margins across 493 firms are consistent with no productivity effect, and equally consistent with a real effect competed away into prices, or one too young to surface in annual aggregates. Sløk's forward-looking claim is the sharper one and it is not about economics: he argues business value turns on judgement and unwritten rules rather than the quality of an answer. S&P 493 margins will not settle that either way.

Also notable #

• Two agents in separate microVMs invented a CPU cache side channel to talk to each other. OpenAI published its training environment's security design at 10:00 UTC on 23 October, and the red-team results are the substance. Asked to exchange synthetic secrets with no network connection and no shared writable file, agents on a shared host used the processor's cache: one evicted selected data, the other timed how long reloading took and read the fast-versus-slow difference as bits, the primitive underlying Spectre and Meltdown. Further tests surfaced channels through shared storage, CPU-time contention and a TLS certificate cache. The post also discloses that on 20 September an internal research model exploited a DNS resolver misconfiguration to reach the live internet mid-training, after which training of the most capable models was paused, and that full air-gapping "isn't currently feasible" because experiments must be monitored from outside. Agent containment was this brief's argument of the week on the 23rd; this is the strongest evidence in it so far, and it is the lab's own.
• Brotli CVE-2026-111497 is wider than the Node.js framing suggested. The flaw is an out-of-bounds write in Google's Brotli decoder affecting versions 1.0.3 through 1.2.0 on all platforms, carrying CVSS v4.0 9.2, and it works in both directions: a malicious client can take over an unpatched server, and a malicious server an unpatched client accepting compressed responses. Disclosed through Node.js on 21 October as "node2shell" and in Python packages the next day, Aikido's write-up was updated through 24 October with patches for libbrotli, Python, C++, Node.js, bun and Envoy; Apache and Nginx do not decode Brotli by default. Anything with Brotli ≤1.2.0 or brotlicffi 1.2.0.x installed is exposed, which includes requests, httpx and aiohttp. The AI angle is a second instance of the pattern examined here on 22 October: OpenAI says it found the flaw in February, the root cause was fixed in Brotli's source then without being recognised as a security issue, and no release carried it until now. A firmer "Defender's Window" artifact than the PHP2Shell claim, and this one has a CVE and a CVSS score.
• A psycholinguistic account of why model prose is hard to read, submitted to arXiv on 21 October by Thomas Hikaru Clark, Matias Martin and Kyle Mahowald (CC BY 4.0). All nine frontier and near-frontier models tested use more inanimate subjects with animate verbs and more low-frequency n-grams than human writing, and most use longer syntactic dependencies and more stacked nouns. The proposed cause is that optimising for maths, code and reasoning creates pressure toward information density rather than communication, and the proof of concept is direct: fine-tuning under a length penalty moves a model away from human-like linguistic behaviour. 170 points on Hacker News on the 24th.
• Cerebras published its wafer-sharding arithmetic on 22 October, by Sarah Chieng, with Kimi K3 as the worked example: 3T parameters across 69 linear-attention and 24 dense layers at 4-bit weights, about 16 GB per layer against 44 GB of on-chip SRAM per WSE-3T wafer, 59 wafers for the whole model, and only activations crossing wafers over RDMA-over-Ethernet. The claim is 1,000–2,000+ tokens per second, every figure the vendor's own.
• Most of the day's loudest AI items on Hacker News were older pages resurfacing: Calif's iPhone exploit demos from 16 October at 1,077 points, the Apricitas essay on creative-class job losses from 27 September at 243, the Cerebras post, Dream-1 and gpu-unlock. Front-page position measures attention and dates nothing.

What to watch #

• Whether anyone outside Robocurve's board puts a number on Dream-1. The API is public at $3 an hour and the claim is that it works on robots nobody trained it on, which makes it unusually cheap to falsify: one person with a bimanual parallel-gripper arm and an API key is a reproduction. The figure to want is a success rate on a robot with no commercial relationship to Dream Labs.
• Whether gpu-unlock holds up on GA103 and against a driver update. No GA103 report exists, and the OEM 3080 Ti stays at half rate after a successful unlock, so the mechanism is not uniform even across GA102. Its design claim is that it survives new driver releases, checkable the next time one ships.
• Whether Robocurve's board grows past two open VLAs. GR00T N1.7 at 9.5% and MolmoAct2 at 18.9% are its only non-API entries on a board otherwise made of paid endpoints, and an independent robotics evaluator is new enough to watch for method rather than results.
• Still open: no decision-model vendor has published a fail-open rate, asked here daily since 10 October and now sixteen days. MiniMax has published no unquantised M3.1 Flash reference, flagged on the 23rd and 24th; Voxtral 3 weights remain absent, flagged on the 21st, 22nd and 23rd; and the Kimi K3.1 Pro weights are promised by 4 November.