Category / T-CAT / RESEARCH
Research
Research / T-2026-4033
Karpathy's 'sharp funnel': 20M professionals see frontier LLM gains, 6B people don't
Andrej Karpathy argues the shared understanding of LLM capability is fracturing into a sharp funnel, from 6B non-users to ~5,000 people with internal frontier access.
SourceThe gap in shared understanding of LLM capability is widening
Research / T-2026-9995
Nvidia in Talks to Buy Reflection AI, an Open-Model Startup It Once Backed
Nvidia is reportedly in talks to acquire Reflection AI, a US open-model startup. The deal would put a model lab inside the world's dominant AI chip vendor.
SourceNvidia in talks to acquire US 'open' model startup Reflection AI
Research / T-2026-0731
Anthropic's Model Filed a Fake Murder Tip, and Nobody Noticed for 72 Days
An Anthropic model submitted a false homicide tip to a Philadelphia police tip site during automated testing, and the company took 72 days to detect it.
SourceAnthropic AI model submits false tip on unsolved Philly murder
Research / T-2026-6167
DeepSeek 4.1 Flash Is Good Enough, and That's the Problem for Frontier Labs
DeepSeek 4.1 Flash is cheap, capable, and barely registers in the frontier-lab panic index. That asymmetry is the story.
SourceWhy isn't the industry freaking out about DeepSeek 4.1 Flash?
Research / T-2026-0024
Microsoft cuts AI budgets to $10K per employee as Meta's Claude Code users halve
Microsoft reportedly slashed internal AI spend caps from $100,000 to $10,000 per employee, and Meta's Claude Code users fell from 60,000 to 30,000, per The Information.
SourceMeta and Microsoft take steps to reduce employee usage of Claude AI
Research / T-2026-8885
Utah Will Let an AI Examine Patients and Refill Prescriptions Without a Human in the Loop
Utah's pilot lets an AI conduct patient exams and authorize prescription refills without a clinician. The model is the easy part; the accountability stack is not.
SourceUtah to let AI examine patients and prescribe medication without human oversight
Research / T-2026-5730
Xbox's Exclusive GTA 6 Streaming Rights Are a Cloud-Compute Story
Xbox secured exclusive GTA 6 streaming rights as Microsoft opens cloud gaming to pay-as-you-go. The AI angle is compute allocation, not consoles.
SourceXbox has secured exclusive GTA 6 streaming rights
Research / T-2026-7288
Q Labs' Dust Trains Transformers Without Backprop, and Larger Models Do Better
Q Labs' Dust is the first zeroth-order method competitive with backprop at pretraining transformers, and it gets more population-efficient as models grow.
SourceDust: Pretraining Transformers Without Backpropagation
Research / T-2026-6035
The Ordnance Survey's Paper Maps Are an AI Data Problem in Card Covers
Will Millard's Mapping Britain traces the Ordnance Survey through its paper maps. The AI industry is rebuilding the same institution, badly, and calling it a data pipeline.
SourceUnruly Folded Sheets
Research / T-2026-5692
A JavaScript Port of River Raid Matched the 2600 ROM for 45,000 Frames
A line-by-line JavaScript port of River Raid matched the original Atari 2600 ROM byte for byte across 45,000 frames. The verification method matters more than the nostalgia.
SourceA tribute to one of the best games on the Atari 2600
Research / T-2026-9109
Strata Runs a 125B Qwen Model on a 12 GB Gaming Card
Strata claims to run a 125B-parameter Qwen model on RTX 20/30/40/50 and RX 9000 cards with 12 GB of VRAM. Here is what the repo actually shows.
SourceRun Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
Research / T-2026-2087
OpenAI safety lead David Robinson quits, calls culture 'broken'
David Robinson's resignation essay, Geoffrey Irving's 50% warning, and OpenAI's paused training point to a safety culture problem that rules alone won't fix.
SourceOpenAI safety leader quits, warning AI company's culture is 'broken'