Category / T-CAT / RESEARCH
Research
Research / T-2026-6858
GLM 5.2 and the coming AI margin collapse
An open-weight model from Z.ai matches frontier quality at 20% of the price. The AI industry's business model relies on inference margins that may not survive.
SourceGLM 5.2 and the coming AI margin collapse
Research / T-2026-4048
The AI industry is learning the wrong lessons from Vasily Grossman
A Baffler essay on Vasily Grossman's Soviet epic carries an uncomfortable mirror for AI builders who mistake scale for meaning.
SourceThe Music of Destruction
Research / T-2026-2278
GPT-5.6 Sol Ultra lands in Codex, not as a standalone model
Tibo Thottiaux confirms GPT-5.6 Sol Ultra will ship inside Codex, not ChatGPT. A bet on agentic coding over general chat.
SourceGPT-5.6 Sol Ultra will be in Codex
Research / T-2026-3820
GPT-5.5 Codex has a reasoning-token clustering problem at 516, 1034, and 1552
An analysis of 390,000 Codex responses shows GPT-5.5 clusters at fixed reasoning-token counts, coinciding with a sharp drop in reasoning intensity and task quality.
SourceGPT-5.5 Codex reasoning-token clustering may be leading to degraded performance
Research / T-2026-5524
Wafer hits 2626 tok/s on GLM5.2 with AMD MI355X at over 2x lower cost than Blackwell
Wafer serves GLM5.2 on AMD MI355X at 2626 tok/s/node and 213 tok/s single stream, outperforming Blackwell on cost while requiring no custom kernels.
SourceGLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than Blackwell
Research / T-2026-0687
Claude-real-video: the hack that lets any LLM actually watch a video
A new open-source tool, claude-real-video, extracts meaningful frames from any video and feeds them directly to an LLM — no cloud upload, no transcript-only blind spot.
SourceClaude-real-video － any LLM can watch a video
Research / T-2026-7117
Parsewise (YC P25) bets that document AI's bottleneck is trust, not extraction
Parsewise, founded by ex-Palantir and Bain engineers, launches an API for cross-document data extraction with word-level citations, beating Gemini on the Databricks OfficeQA…
SourceLaunch HN: Parsewise (YC P25) – Reason Across Documents with an API
Research / T-2026-0414
Meituan's LongCat-2.0: A 1.6T-Parameter MoE Trained on Chinese Chips, Now Open Source
Meituan releases LongCat-2.0, a 1.6T-parameter open-source MoE model trained on 50,000 domestic chips, with a 1M context window and top-3 OpenRouter rankings.
SourceLongCat-2.0, a large-scale MoE model with 1.6T total and 48B Active
Research / T-2026-8377
Central bankers warn debt-fuelled AI boom risks a global financial crash
The Bank for International Settlements warns that debt-fuelled AI spending, opaque financing, and shadow bank lending risk a global financial crash similar to the 2008 credit…
SourceAI boom risks global financial crash, warn central bankers
Research / T-2026-6838
The single Unicode character that reveals AI's typeface blind spot
A deep dive into Arabic ligature rendering exposes a blind spot in how AI systems understand written language.
SourceDeciphering Basmala
Research / T-2026-1811
DeepSeek's DSpark spec-decoding framework accelerates inference 60-85%
DeepSeek open-sources DeepSpec, a full-stack speculative decoding library, claiming 60-85% speedups on Flash models and 57-78% on Pro models.
SourceDSpark: Speculative decoding accelerates LLM inference [pdf]
Research / T-2026-9898
Marfa Public Radio's Sleep Podcast Is a Perfect Test for AI Listening
A small West Texas public radio station's fundraising gimmick — reading boring documents aloud — accidentally creates a perfect dataset for studying how AI models handle long…
SourceMarfa Public Radio Puts You to Sleep