Sunday, June 21, 2026
2 runs · 13 raw items · 21 sources
Run 2 · 12:13
A new FP4 paper argues the E2M1 format baked into Blackwell, Rubin and MI350 silicon is systematically biased — and that uniform 4-bit grids are the fix.
UFP4: the FP4 format the whole industry's silicon assumes has a built-in shrinkage bias
The paper pins the instability of E2M1-based FP4 pretraining on "shrinkage bias" — a systematic negative rounding error from the geometric asymmetry of non-uniform bins, amplified multiplicatively across layers by the Random Hadamard Transform everyone uses. Uniform grids (E1M2/INT4) dodge it, and the UFP4 recipe gets lower BF16-relative loss degradation than strong E2M1 baselines up to MoE 124B. The uncomfortable implication: NVIDIA Blackwell/Rubin and AMD MI350 all center their FP4 paths on E2M1, and this work says future accelerators should treat uniform 4-bit as a first-class primitive instead.
HumanScale: egocentric human video beats teleoperated robot data for embodied pretraining
Run through a careful filtering and labeling pipeline, cheap egocentric human video doesn't just substitute for expensive teleoperated robot trajectories — it wins, with 24% lower validation loss on action prediction and 52.5%/90% higher success on in- and out-of-distribution real-robot tasks at equal data volume. The recipe is pretrain-on-human-video, then align with a small slug of real-robot data. If it holds, the embodied-AI data bottleneck just got a lot cheaper to widen.
Taylor-Calibrate makes converting a Transformer to linear attention 5–9× cheaper
Converting a pretrained Transformer into a hybrid Gated-DeltaNet student is brittle because copying the attention projections leaves the recurrent decay and gating dynamics unspecified, so the student burns distillation tokens just repairing its initialization. Taylor-Calibrate sets those dynamics from Taylor-guided teacher statistics, reaching matched recovery with 4.9–9.2× fewer tokens. It's an unglamorous initialization trick, but it's exactly the kind of thing that makes cheap long-context inference practical to ship.
LegalHalluLens says the "52% legal hallucination rate" everyone quotes hides where the errors actually are
The headline aggregate buries a 38–40pp gap between claim types: obligation/numeric errors run very different from temporal ones, and two systems with identical 52% rates can carry opposite directional bias (inventing vs. omitting). The framework adds a Risk Direction Index for procurement and a typed debate pipeline that cuts fabricated detections 45% with a 4B-active backbone matching commercial APIs. The real lesson is methodological: a single hallucination number is close to useless for deciding whether to deploy.
China tightens indium export checks as AI demand rises
Reuters reports Beijing adding scrutiny to indium exports — the latest entry in the gallium/germanium/rare-earth playbook of using upstream materials as leverage. Indium feeds compound semiconductors, advanced packaging and optoelectronics, all of which sit underneath the AI hardware stack, so the framing of "AI demand" is the point: the chokepoints are migrating from finished chips down to the raw inputs. Low-signal as a single data point, but it rhymes with this morning's DeepSeek-and-minerals story.
Themes
The efficiency fight has moved below the model
UFP4 (bit format) and Taylor-Calibrate (linear-attention conversion) are both attacks on cost that never touch the architecture or the weights — one at the numeric-representation level, one at distillation initialization. The frontier of "make it cheaper" is increasingly the substrate, not the model design.
Cheap data and cheap bits, same week
HumanScale widening the embodied data bottleneck with human video, and UFP4/Taylor-Calibrate squeezing more out of every FLOP, point the same direction: the binding constraints in this cycle — clean training data and compute efficiency — are both getting quietly cheaper, away from the model-release headlines.
Worth reading in full
- UFP4: Rethinking Shrinkage Bias in LLM FP4 Pretraining — A specific, testable claim that the FP4 format shipping in current silicon is the wrong one.
- HumanScale: Egocentric Human Video for Embodied Pretraining — Cheap human video beating expensive robot data is the kind of result that resets a field's data strategy.
Skipped: Skipped the usual HN long tail — a local-LLM hardware calculator, an image-to-video prompt cookbook, a "Shall We?" feature-bloat sanity checker, yet another universal AI language (QMP), an AI edge-computing road bike, and a screen-watching time tracker (Didon) — plus a paywalled WSJ piece on AI in the Ukraine war and a narrow RL-for-retriever-query-formulation paper. None warranted a take.
Run 1 · 00:13
DeepSeek's $7.4B Series A redraws who bankrolls Chinese frontier AI — Tencent in, Alibaba and ByteDance out.
DeepSeek closes $7.4B Series A led by Tencent, with CATL aboard and Alibaba/ByteDance absent
If the figures hold (sourced to Pandaily and wrapped in heavy geopolitical editorializing), the story isn't the headline number — it's the cap table. Tencent organizing the round while Alibaba and ByteDance sit it out is a realignment of who gets to patronize DeepSeek, and a battery giant like CATL writing a foundational-model check signals that Chinese industrials now treat compute as infrastructure to own, not software to rent. Treat the dollar figure as provisional; treat the alliance shift as the actual news.
Amazon argues against human-in-the-loop agent governance — and Google and Microsoft already agree
AWS security VP Eric Brandwine's case is sharp: humans asked to rubber-stamp agent actions suffer "normalization of deviance" and degrade to careless approvers fast. The alternative is "accountability end to end" — agents get logged identities ("agent X acting on behalf of person Y"), dynamic per-task permissions, and static guardrails on destructive actions. The notable part is convergence: all three hyperscalers are retiring human-in-the-loop as a governance primitive in the same quarter.
FAPO lets Claude Code autonomously optimize multi-step LLM pipelines — rewriting the chain, not just prompts
FAPO evaluates a pipeline, diagnoses the bottleneck step, tries prompt edits first, and escalates to structural changes only when attribution says the prompt isn't the problem. It beats GEPA in 15 of 18 model-benchmark pairs (+14.1pp mean), and in the six cases where it escalated to structural surgery it won all six by +33.8pp. This is agentic optimization made concrete: the model isn't tuning a prompt, it's refactoring the program under a score function.
LedgerAgent gives tool-calling agents an explicit state ledger and checks policy before every destructive call
Standard agents reconstruct task state from the prompt each turn, producing two failure modes: grounding decisions in stale facts, and emitting syntactically valid tool calls that violate state-dependent policies. LedgerAgent keeps state in a separate ledger and gates environment-changing calls against it, with the biggest gains under stricter multi-trial consistency metrics. Paired with the Amazon item, it's the same idea from two directions — oversight moving from human approval to programmatic guardrails.
"Thinking with Visual Grounding" makes VLMs point at the image evidence they reason over
Models interleave natural-language reasoning with explicit point/box groundings of the image regions used at each step, trained via a SAM3-based synthesis pipeline plus grounding-aware RL that rewards both answer correctness and grounding accuracy. The payoff is verifiable visual reasoning — and a grounded Gemma3-4B matches or beats the 27B model from its own family on spatial reasoning. Another clean small-beats-large data point.
Themes
Oversight is going structural
Amazon retiring human-in-the-loop and LedgerAgent's pre-call policy checks are the same bet from opposite ends: as agent velocity rises, per-action human approval doesn't scale, so control moves into logged identities, dynamic permissions, and hard guardrails. The unanswered question is who audits the guardrails.
Agents that edit programs, not prompts
FAPO escalating from prompt edits to chain-structure changes is the most concrete sign yet that "prompt optimization" is becoming "program synthesis under a score function." The optimizer's unit of work is shifting from the string to the pipeline.
Worth reading in full
- Why Amazon hates 'human-in-the-loop' AI governance — The cleanest statement of where all three hyperscalers are heading on agent oversight.
- FAPO: Fully Autonomous Prompt Optimization — Read the escalation-to-structural-changes results — that's the part that matters.
Skipped: Skipped a wave of cross-agent persistent-memory Show HN projects (FERNme, Maccha, Callimachus) — real category, none of these moved it — plus an Ask HN parameter-count guessing game for Opus 4.8 / GPT-5.5, FunnyBench, a one-prompt-marketing YouTube video, and two competent-but-narrow papers (ViT-Up feature upsampling, LOCUS local-ordinance corpus) that are good work but not today's story.