Apollo sees no broad AI payoff outside tech, but its charts cannot isolate AI’s effect
Apollo’s comparison of seven technology giants with the other 493 S&P 500 companies finds no broad financial lift outside the giants, without establishing AI’s causal effect.
RTX compute unlocks reach Ada and Turing, with gains narrower than the 2× headlines
Separate open-source projects report faster selected tensor operations on named RTX 20, 30 and 40 cards, without establishing a universal doubling of chatbot speed.
Lean-QwenFlash reports a 125B model at 20 tokens a second with 16 GB of GPU memory
Lean-QwenFlash streams model experts from solid-state storage and reports 20 tokens a second at a 262,000-token context, with quality retention still unmeasured.
ISTA’s two-bit DeepSeek quant matches three reported tests in a 249 GB package
ISTA-DASLab’s compact DeepSeek release matches its base model on three reported benchmarks, while requiring a 248.84 GB download and substantial memory outside the GPU.
Forced-choice election audit finds opposite party skews in GPT and Grok API answers
Convovo’s search-enabled election audit finds strongly opposite party skews among candidate-naming answers from GPT and Grok, with refusals excluded from those percentages.
Reported Claude TV-repair refusal exposes the gap between permitted research and safeguards
A first-person account says Claude interrupted an attempt to restore root access on an owned LG television, while Anthropic’s forthcoming policy permits lawful owner-authorized research.
A community hub starts turning OpenAI’s math manuscripts into reviewable work
A community revision hub announced on Terence Tao’s blog organizes corrections, citations and reviews for a current catalogue of 719 OpenAI mathematical manuscripts.
1Password launches agent logins that keep passwords outside the model
1Password for Agents brokers approved sign-ins without showing credentials to the model, while leaving the logged-in agent able to act inside an account.
AI prose study finds dense language patterns, without testing reader comprehension
A new study finds that many AI outputs overuse language features associated with processing difficulty, but it does not measure how quickly or accurately people read them.
A small teaching study finds confident AI falsehoods survive repeated questioning
A 32-person experiment found that learners often absorbed planted AI falsehoods while remaining confident, under instructions to rely only on the assistant.
ETH researchers use 13.2 minutes of human motion to guide humanoid contacts
ETH researchers report a humanoid mobility system that turns 13.2 minutes of human motion into contact-planning guidance, rather than arbitrary task learning.
Dream Labs pairs a $90 million seed round with a cross-robot action API
Dream Labs has announced a $90 million seed round and a public Dream-1 API that uses shared end-effector goals to attempt control across different robots.