Category / T-CAT / RESEARCH
Research
Research / T-2026-8237
Apple's Gemini Pivot: A New AI Architecture, and a Quiet Admission
Apple reveals a major overhaul of Apple Intelligence, co-developing foundation models with Google and centering them on Gemini technology.
SourceApple reveals new AI architecture built around Google Gemini models
Research / T-2026-2092
Mach's self-hosting milestone: a systems language built for AI's infrastructure moment
Mach, a compiled systems language that eschews LLVM and libc, reaches self-hosting. Its anti-magic philosophy and explicit design have implications for AI infrastructure builders.
SourceShow HN: Mach – A compiled systems language looking for contributions
Research / T-2026-3605
DeepSeek V4 Pro beats GPT-5.5 Pro on precision — and on cost, the gap is brutal
DeepSeek V4 Pro beats GPT-5.5 Pro on precision in a fresh benchmark, but the real story is cost: roughly 1/200th the price per task.
SourceDeepSeek V4 Pro beats GPT-5.5 Pro on precision
Research / T-2026-4674
Meta confirms 20,000-plus Instagram accounts hacked by abusing its AI chatbot
Meta confirms 20,225 Instagram accounts were hijacked by abusing its AI chatbot. The bug let attackers reset passwords without proper email verification.
SourceMeta confirms 1000s of Instagram accounts were hacked by abusing its AI chatbot
Research / T-2026-3941
General Instinct shrinks a 245 GB model to fit on a single GPU
YC P26 startup General Instinct open-sources InstinctRazor, compressing a 245 GB MoE model into 48 GiB while beating Google's Gemma-4-26B on MMLU-Pro and GPQA-D.
SourceLaunch HN: General Instinct (YC P26) – Frontier models on edge devices
Research / T-2026-6217
Anthropic's internal data shows AI is closing the loop on its own development
Anthropic releases internal data on recursive self-improvement. Engineers ship 8x more code. Claude writes 80% of merged code. The path to autonomous AI development is visible.
SourceWhen AI Builds Itself: Our progress toward recursive self-improvement
Research / T-2026-1742
Anthropic details how it contains Claude across three products
Anthropic's engineering blog on containing Claude across claude.ai, Claude Code, and Cowork reveals the tradeoffs between capability and blast radius.
SourceThe ways we contain Claude across products
Research / T-2026-5798
Law Professors Prefer AI Answers to Their Peers' in Stanford Study
A new Stanford Law study finds professors overwhelmingly prefer AI-generated answers to student questions over responses written by fellow instructors, with implications for…
SourceAI outperforms law professors in Stanford Law study
Research / T-2026-4360
Alphabet's $80B equity raise signals a compute crisis, not a capex plan
Alphabet plans to raise $80B through stock sales, including $10B from Berkshire Hathaway, to fund AI compute infrastructure. The move signals demand far outstripping supply, with…
SourceAlphabet announces $80B equity capital raise to expand AI infra and compute
Research / T-2026-6338
4x faster, thinking bigger: what AI prototyping actually changes
Daryl Cecile reports 4x faster PRs after a year with AI agents. The real shift is in how he plans work, not just how fast he types.
SourceThe Speed of Prototyping in the Age of AI
Research / T-2026-1602
Open Envelope wants to be the Dockerfile for AI agent teams
Open Envelope ships an Apache 2.0 schema for defining AI agent teams with roles, handoffs, and network-level access controls. The ambition is a portable format. The challenge is
SourceShow HN: Open Envelope – an open schema for defining AI agent teams
Research / T-2026-0356
Liquid AI's 8B-A1B MoE: 38T Tokens, 128K Context, and a Bet on Reasoning
Liquid AI's new 8B-A1B MoE, trained on 38T tokens with a 128K context window, challenges assumptions about on-device model capability.
SourceLiquid AI reveals 8B-A1B MoE trained on 38T