Robots need a testbed before they need more data
Ai2's ToyBlocks gives robot learning a controlled world with 85,336 demonstrations. A compact policy beats generative backbones. And a new method infers hidden interiors from collisions.
Updated 2026-11-07 12:00 UTC
26 stories, 114 Manifests cited.
Artificial Intelligence
RL forgets, adds little over good fine-tuning — and still gets better at the hard parts
Four papers and one analysis put reinforcement learning under the microscope: it forgets old skills, it barely beats good supervised fine-tuning on one benchmark, and new methods for rubrics and agent training push it forward anyway.
The refusal test: the same biology question gets refused, answered, or ignored depending on the model
Benchling's new benchmark finds open-weight models rarely refuse hazardous biology requests. A chemistry audit finds the benchmarks themselves are confounded. And a post-training run shows values are one fine-tune away.
Read, don't copy: the data layer stops moving tables around
Databricks wants teams to read shared tables in place. Snowflake can now query Databricks tables directly. The week's data news is that the copy is the problem.
The new tests grade agents on real work — and the scores are humbling
FrontierCUA's best model passes 25% of enterprise workflows. OpenFrontiers finds closed-book rediscovery is rare. And a math modeling bench checks answers against Lean 4 proofs.
Nobody wants one model anymore: routers and multi-model habits are the hedge against lock-in
Heavy AI users already spread work across a median of six model families. OpenRouter's rebuilt router picks the model per request. And a new framework warns that even cross-model collaboration may be fooling itself.
Legal AI stops being a chatbot and becomes the team
GC AI now serves more than 2,200 corporate legal departments and beats ChatGPT-5.5 on a legal benchmark. GenieAI's specialist agents pass 11 of 11 contract checks. And Harvey shows the real work is renewals.
The forecasting audit: random splits lie, live benchmarks bite, and operators still win
A steel model that looks great on random splits collapses on new steel families. An energy benchmark run live shows grid operators still beat foundation models. And a volatility method improves every comparison it was tested on.
Diffusion language models get their own training stack — RL, correction, and faster decoding
Five preprints in one week give diffusion language models the pieces autoregressive models already had: a way to do RL, a way to fix their own mistakes, and a way to decode in parallel without losing accuracy.
Inference is not limited by FLOPs. It is limited by memory bandwidth.
A 320B mixture-of-experts model needs 18 GB of weight movement per token. Testing every serving configuration can take 25,000 GPU-hours. The week's inference news is about the bottleneck, not the model.
Brain decoding gets cheaper: behavior-only training and foundation models that skip fine-tuning
A motor-decoding framework trains on behavior alone. An EEG foundation model needs only linear probing. And sleep staging hits 0.967 AUC. The brain-signal stack is getting the foundation-model treatment.
How much should AI do on its own? The answer is a dial, not a number
Google found coding agents get distracted by TODO comments in 913,000 sessions. GenieAI runs agents on escalation rules. Zenity, Aria, and SAS all say the same thing: design the stop, not just the go.
Anthropic's S-1 says the lab is accidentally profitable. The market just lost $2.5 trillion.
Revenue grew 14x while serving compute grew 2.3x. The filing reframes the AI trade around efficiency, just as tech stocks had their worst day since June.
Agents are learning to remember — by dreaming, compressing, and governing what they keep
Anthropic lets agents reorganize memory in idle time. A new method compresses tool output into latent tokens. And enterprises are learning that the context layer needs governance, not just retrieval.
Finance & Investing
This quarter's cash went back to owners, not into growth
Franklin Resources, Plains All American Pipeline, Lifetime Brands, Calumet and Global Partners all sent cash to shareholders or creditors. The message is about what they did not spend it on.
Crypto got two doors this week: one opened, one shut
The Bank of Russia's retail opening and OFAC's designations under Operation Economic Outcast show the same asset being pulled into the regulated system and pushed out of the illicit one.
Capital didn't leave this quarter. It moved.
The quarter's filings show money rotating toward AI-linked demand and alternatives, and away from consumer spending and legacy managers.
Bonds won the election week, and the ballot box did half the work
A slower AI buildout means less corporate bond issuance, weaker growth and lower inflation — and the market noticed before the week was out.
The Bank of England hiked into the selloff. The Fed may be next.
An 8-1 vote for a half-point rise, a jobs report that beat forecasts, and a dollar that rises with every equity fall — the rate story has flipped from cuts to hikes.
The AI correction was already underway before the election made it worse
The midterm results accelerated a selloff that started earlier, and the market's answer so far is to favour AI capacity that already exists over capacity still to be built.
Business Podcasts
Anthropic's filing gave podcasts one set of numbers and four arguments
One lab's efficiency gains, a debated $1.2–2 trillion valuation, a warning about moats and a new word for the selloff: FOGO.
Health & Biotech
Heart risk is being found earlier, measured cheaper, and still arriving late
A GLP-1 study, a national burden analysis, a blood-count marker and a heart-failure staging test agree: the risk is visible early. The system still acts late.
Sports
Baseball's offseason started in the front office, not on the field
A $23 million extension in Washington, a new GM in St. Louis, a hitting coach out in Los Angeles and two staffs rebuilt: the week's moves were about who decides.
Federal Government
Washington's enforcement week, in four dollar figures
A $73 million crime grant, a record $49.75 million landlord settlement, a $1.25 million kickback case and an $80.5 million drug seizure landed in the same week.
State Legislatures & Executives
The election is over. Statehouses are already building 2027
Minnesota picked its House leaders, New York moved an AI liability bill, Arkansas proposed a $6.97 billion budget and South Dakota began its transition.
Energy & Climate
The satellites watching Earth keep arriving with their own error bars
From Himalayan slopes to the Mekong delta, the new instruments work. The honest part is how precisely their limits are now stated.