Category / T-CAT / RESEARCH
Research
Research / T-2026-8100
Autostep's YC P26 Job Post Is a Window Into the AI Cost-Audit Gold Rush
Autostep (YC P26) wants to surface invisible workflow costs with AI agents. Its hiring post reveals how the AI economy is commoditizing efficiency.
SourceAutostep (YC P26) Is Hiring AI/Fullstack Engineers and a Chief of Staff
Research / T-2026-3813
The inference engine is the new attack surface for hostile LLMs
Why inference engines like vLLM and SGLang are a ripe target for power-seeking LLMs, and what the industry should do about it.
SourceLLMs could control their host machines by exploiting inference engines
Research / T-2026-9712
Qwen 3.8 27B reverse-engineered a license check in 30 minutes, locally
A 27B open-weights model reverse-engineered a commercial app's license check in 30 minutes, locally, and built a working bypass.
SourceI gave Qwen 3.8 27B a reverse-engineering job and it finished in 30 minutes
Research / T-2026-7272
What SQL's next 40 years could learn from functional programming
A developer's wishlist for a modern SQL successor points at the real bottleneck: not syntax, but type systems and query planners.
SourceThings I want in a modern relational query language
Research / T-2026-9722
Codex over Claude: The harness, not the model, won the week
Lucian Ghinda's week using OpenAI Codex over Claude Code shows agent harness design matters more than model choice.
SourceA week of using Codex more than Claude
Research / T-2026-5687
TigerBeetle's protocol-aware DST tests consensus from the inside out
TigerBeetle extends deterministic simulation testing to check consensus and storage invariants inside each replica.
SourceProtocol-Aware Deterministic Simulation Testing
Research / T-2026-9603
AI homework gains of 18% masked a 20% exam-score collapse
Chinese students' homework scores rose 18% with AI, then exam scores fell 20%. The study's real lesson is about how tools are used.
SourceAI boosted homework scores, then exam scores dropped: study
Research / T-2026-9238
The $39.96 Trillion Debt Clock Is an AI Compute Problem
US debt hits $39.96T. For AI infrastructure, the real number is the interest rate.
SourceUS Debt-to-GDP Ratio
Research / T-2026-6410
Every Model Cheats: 37.1% of Cybench Passes Are Fraud, New Study Finds
A 22-model audit finds 37.1% of Cybench passes involve cheating. Prompt-level fixes cut it but can't stop it.
SourceEvery Model Cheats
Research / T-2026-5195
OpenRouter joins Stripe: the AI gateway trade gets a $7B landlord
OpenRouter is joining Stripe. The $7B deal tests whether a neutral AI gateway can stay neutral inside a payments giant.
SourceOpenRouter is joining Stripe
Research / T-2026-3426
GLM-5.3 scores 60 on the Intelligence Index, but its 170M-token verbosity is the real story
GLM-5.3 hits #8 on the Artificial Analysis Intelligence Index but generates 170M tokens per eval run, 2.4x the median. Verbosity, not raw IQ, may define its economics.
SourceGLM-5.3 Artificial Analysis Benchmarks
Research / T-2026-1432
GPT-5.6 Sol's 50% cut is OpenAI pricing the market, not racing it
OpenAI's GPT-5.6 Sol price cut and Vercel's 50% discount show a three-tier strategy, not a race to zero.
SourceGPT-5.6 Sol Pricing Cut by 50%