AI Tech Daily - 2026-11-15
2026-11-15
| 2026-11-15
字数 1916阅读时长≈ 5 分钟

📊 Today's Overview

The open-source AI debate got a sharp new framework today, as Nathan Lambert sketched what a post-Jensen Nvidia means for open weights. Meanwhile, Microsoft researchers showed a backdoor can be planted by rewriting fewer than 10 neurons — and standard evals stay green. On the hardware side, Google Cloud's eighth-gen TPUs split into 8tx (first with confidential computing) and 8i Max (dual-die packaging), while H100 rental prices across clouds now span a 2.8x gap. And CoRL 2026 drew 2,094 submissions, with SuperMimic training humanoid motion tracking on a single H100 in ~3 hours.

🔥 Trend Insights

  • Efficiency over raw scale: SuperMimic hits humanoid motion tracking in ~3 H100 hours versus NVIDIA's ~21,000 GPU hours, and NVIDIA's Dory1.0-4B reuses 20 shared layers to beat bigger models — architecture is doing the heavy lifting.
  • Model weights as audit surface: Researchers claim GPU fingerprinting can infer which hardware trained DeepSeek and Kimi, while others show training history — data order, LR curves, seeds — is readable straight from weights.
  • Agent reliability gets measured: Garry Tan's g-harness-evals finds five agent frameworks clustered at 71–79% pass rates with 2x cost gaps, and a new checklist pushes pass^k consistency over one-off demos.

🐦 X/Twitter Highlights

📈 Trends & Buzz

  • Anthropic's pre-training compute-efficiency curve called "illustrative only" — The chart's log-scale y-axis shows algorithmic efficiency rising super-exponentially; former Anthropic employee @hilbertspaess says it was a major factor in his departure. Omar Khattab notes the tiny footnote says data points are placed on the chart "for illustration only" @lateinteraction (MIT professor / DSPy author) | @hilbertspaess (former Anthropic employee who says the chart drove his exit)
  • ASUS unveils Gallop, its first factory humanoid robot — It learns manufacturing tasks by watching skilled workers and practicing on real production-line parts; the company plans customer pilots in 2027 @roboticomarket (robotics content account)
  • Panasonic forms ~250-person Physical AI R&D team — Focused on 3D spatial perception, precision robot control, and multimodal AI, with the goal of teaching AI the judgment veteran technicians apply to positioning, friction, and assembly force; validation in its own factories over the next 3–5 years, then warehousing and logistics @CyberRobooo (robotics blogger)
  • vangrid crowdsources ground-level 3D data via phones — Has accumulated 1.52M verified collections, ~490K active nodes, and $740K in USDC bounty payouts; faces and license plates are blurred on-device, data is attested with Merkle trees on Base, and buyers escrow funds before work begins @carte_free (crypto / physical AI observer)

🔧 Tools & Products

  • YC president Garry Tan open-sources agent harness eval g-harness-evals — Fixes the model and tasks, swapping only five frameworks: Claude Code, Codex, OpenClaw, Hermes Agent, and QM. All five land at 71%–79% pass rates on 20 dev tasks, with overlapping 95% confidence intervals. Cost gaps are large: $0.44 per task for OpenClaw, $0.22 for Claude Code, $0.17 for Codex. First-request context handed to the model was roughly 12,900, 18,416, and 20,613 tokens respectively. His own skill pack gstack was executed by the model 0 times across 46 tests, doubling cost instead; even after adding an interception hook, Opus still bypassed it 40 times by writing source directly via Bash. When the team's shared knowledge base gbrain lacked permission isolation, OpenClaw and Hermes leaked 123 and 121 unauthorized records respectively @garrytan (President & CEO, Y Combinator)
  • OpenAI details its open-source Codex harness and code mode — Jeff Chan and Nikunj Handa (OpenAI staff) discuss developer-side progress on the Supabase podcast: open-sourcing the Codex shell, code mode, training models to proactively use async tools, expanding the Responses API, and arguing models should be judged on cost per task rather than price per token @supabase (open-source backend platform)

⚙️ Engineering Practice

  • CoRL 2026 recap: the bottleneck shifts from policy architecture to data and compute economics — 2,094 submissions (~3x last year), 687 accepted (32.8%). Best paper SuperMimic (UC Berkeley) learns a differentiable world model covering an entire motion library, completing humanoid motion tracking on a single H100 in ~3 hours versus NVIDIA SONIC's ~21,000 GPU hours; it failed on 25 of 129,514 motions versus 179 for the baseline, and has been transferred to the Unitree G1. NVIDIA and UCLA's GRAIL generates 20K+ humanoid mobile manipulation sequences from 3D assets plus video priors, hitting 84% grasp success and 90% stair-climbing on the G1. DreamZero is a 14B world-action model running at 7 Hz that adapts to a new body with ~30 minutes of play data @yoshi_flux (world model researcher, WorldFlux)
  • NVIDIA releases Dory1.0-4B, a recurrent Transformer — Reuses 20 shared layers run 1–4 times, so parameters don't grow with depth; SWE-bench Verified rises from 50.3 at 1 pass to 59.5 at 2 and 61.8 at 4, and the 12-benchmark average goes from 35.17 to 40.60, surpassing Qwen3.5-9B, Gemma 4 12B, and 26B-A4B at 2+ passes. End-to-end compute grows only 1.03–1.25x, and 2 passes is the recommended default. Limitation: on AA-Omniscience it stays at 10.0 at both 1 and 4 passes, still trailing Qwen3.5-4B on knowledge tasks @_avichawla (AI tutorial author, DailyDoseofDS)
  • SatBench: 63% of 173 cited benchmarks since 2023 are saturated — Ethan Mollick tallied the benchmarks cited in papers from various labs and measured how fast they get "cracked," concluding saturation is accelerating year over year @emollick (Wharton professor)
  • Reading a model's training history straight from its weights — Alapan Chaudhuri lists what can be recovered from weights: tokens in the data, learning-rate curves, step counts, shuffle seeds, context length. He notes OLMo 2's first instruct version had 20,341 embedding rows that should have trained but never moved, including the end-of-turn marker; AI2 later found the tokenizer mismatch and retrained. Pythia-1B's actual peak LR was 2.5e-4 versus 3.0e-4 on the model card, and the LR was flat for the last 28,000 steps. Qwen2.5 instruction tuning moved only 4 of 22 control tokens, identically across 7B/14B/32B/72B. Christopher Potts adds that logits alone can recover training-data order and reveal test-set contamination @alapanchaudhuri (independent researcher) | @ChrisGPotts (Stanford linguistics professor)

⭐ Featured Content

Nathan Lambert sketches "open-source AI without Jensen": Nvidia's financial motive remains, but founder-style funding could dry up | An upside/downside framework for the open-model economy
The first sober take on where the open ecosystem goes after Jensen Huang steps down. The core is one pro, one con: on the plus side, open source has locked up two lanes — ownership of intelligence and enterprise post-training. On the minus side, with compute extremely scarce, open weights are far from commoditized (DeepSeek DeepInfer's high inference margins are the counterexample), and closed labs gain the edge through scale and model-datacenter-chip integration. The author also flags that the real near-term risk isn't technical but clumsy state-level regulation and "open-weight hawks," plus a media ecosystem driven by timing rather than information. Worth it if you want a reusable framework for judging open vs. closed.
Microsoft Research: rewriting fewer than 10 neurons in an LLM can plant a "sleeper backdoor" that passes all standard evals | Backdoor cost pushed down to single-digit neurons
It drops the cost of planting a backdoor from retraining/fine-tuning to a handful of neurons — the model goes all-green on standard evals yet produces attacker-chosen behavior under specific triggers. This poses a new challenge for model supply-chain audits and pre-release weight integrity checks: behavior-based safety screening may fail entirely. Details (trigger mechanism, cross-model transferability, detection methods) weren't fully captured; check the original paper.
Sources: csoonline.com
"GPU fingerprinting": statistical inference on open model weights to identify training hardware, pointing at DeepSeek/Kimi using restricted Blackwell | Training compute provenance becomes externally auditable from the model itself
Researchers claim statistical analysis of open model weights can detect correlation with specific Nvidia hardware, inferring that Chinese open models like DeepSeek and Kimi may have trained on export-controlled Blackwell chips. If the method holds, training compute provenance becomes externally auditable from the model itself, with direct implications for export-control enforcement and open-model compliance disclosure. But the captured content is mostly page scaffolding — method, sample size, false-positive rate, and the named model list all need verification in the original. Treat as a lead, not a conclusion.
Google Cloud's eighth-gen TPU: two new systems — 8tx is the first with confidential computing, 8i Max doubles bandwidth and SRAM via dual-die packaging | Reusing existing chips, changing only packaging/host/network/security
TPU 8tx, built with AMD, encrypts the full host→TPU path plus attestation, protecting both enterprise input data and model weights; TPU 8i Max packages two 8i dies together, doubling bandwidth and on-chip SRAM and adding 33% more HBM capacity. Key insight: neither uses a new accelerator design — only packaging/host/network/security changed. And 8i Max's bandwidth and capacity gains answer different questions: how fast context data moves vs. how much fits. The piece also warns that the 2x latency claims use different baselines and can't be compared directly on cost.
November 2026 cloud GPU pricing comparison: the same H100 runs $12.29 on Azure, $6.88 on AWS, $4.39 on RunPod — a 2.8x spread | A budget sanity check normalized from "per GPU hour" to "per million tokens"
Normalizes hourly list prices for H100/H200/B200/MI300X across providers, noting that 2026 ends years of rental price declines: median on-demand H100 pricing is up 31% year over year, AWS raised Capacity Block prices ~15% in January, and Crusoe is raising prices from December. The methodology (converting instances/GPUs/nodes/contract terms to per-GPU-hour, then to cost per million tokens) is reusable, but the data is mostly secondhand and the site quality is low — use it as a budget reference, not a decision basis.
Bessemer rebuts the "robot bubble" thesis: 745 software companies raised over $30M in the past five years, versus just 42 robotics companies | Firsthand comparison numbers on physical AI private-market sentiment
The Robots & Startups newsletter rounds up three things: the SVR physical AI demo day (Devol Robots won; 3 of 5 finalists do medical robotics, which the author sees as the ideal near-term deployment scenario — semi-structured environments between industrial and home); Bessemer's "18x fewer companies but a 30x bigger market" comparison, framing this as catch-up capital rather than a bubble; and Henrik Christensen's 2026-2035 roadmap, listing physical AI (large-scale VLA training), advanced materials, and next-gen computing as the three enabling trends. Good for quickly calibrating the physical AI capital narrative.
San Francisco startup Foundation aims straight at the Pentagon with humanoid robot Phantom MK-2 | A defense-industry fork in humanoid commercialization
A 200-pound, 6-foot aluminum humanoid, dust- and water-resistant, strong enough to lift an adult, with a physical exclusion zone for "unlimited force" mode. The value isn't in technical detail but in showing a commercialization path distinct from Figure/Tesla — defense orders may deliver paying scenarios and a real data loop sooner than general household chores. A signal worth watching for anyone tracking embodied AI deployment paths and regulatory boundaries, though it lacks the VLA architecture, data plan, and control stack details practitioners actually want.
An agent reliability eval starter checklist: results plus trajectory, 8 failure categories, and τ-bench's pass^k | From "the demo works" to "consistent across repeated trials"
Starts from "a working demo ≠ reliable" and stresses looking at both outcomes and execution traces, distinguishing "legal tool calls" from "task actually completed" (with a customer-service example of canceling the wrong order). It gives 8 failure categories (intent understanding / tool selection / parameter generation / planning / tool execution / context management / error recovery / safety permissions) and uses τ-bench's pass^k to measure consistency across repeated trials. All concepts come from existing papers with no new data or code — a good starting point for an internal team eval checklist.
Sources: dev.to
  • AI
  • Daily
  • Tech Trends
  • AI Tech Daily - 2026-11-16AI Weekly 2026-W46
    Loading...