Category / T-CAT / RESEARCH
Research
Research / T-2026-9325
HarnessTax: The Scaffolding Around a Coding Agent May Matter More Than the Model
A benchmark called HarnessTax measures how much a coding agent's harness, not its model, drives performance. The implication: evals are measuring scaffolding.
SourceHarnessTax: How Much Does the Harness Matter for Coding Agents?
Research / T-2026-4359
Irregular Ran the Evals That Let Claude, GPT, and Meta Models Hack Real Systems
One Israeli EA-linked firm, Irregular, supplied the unsecured environments behind AI hacking incidents at Anthropic, OpenAI, and Meta. The misalignment story does not survive the…
SourceA single firm is behind OpenAI, Anthropic, and Meta hacking scandals
Research / T-2026-7794
Cua's Founding GTM Hire Is a Bet That Computer-Use Agents Are Ready to Sell
Cua is hiring a founding technical GTM lead at $150K-$200K. The role description reveals how computer-use agents are being commercialized.
SourceCua (YC P25) Is Hiring a Founding Technical GTM Lead
Research / T-2026-7259
Lina Khan Says AI CEOs Can Be Charged Under a 1934 Precedent
Lina Khan points to FTC v. R.F. Keppel & Bro to argue frontier labs and their executives can be prosecuted under laws already on the books.
SourceEx-FTC boss Khan: break out the handcuffs for AI CEOs, citing 1934 precedent
Research / T-2026-1969
Anthropic's Hidden PaaS: Inside Claude Code Web's Unstripped Go Binary
Claude Code Web runs on Firecracker MicroVMs restored from frozen snapshots, and ships a 27MB unstripped Go binary that exposes an unreleased Anthropic deployment platform.
SourceReverse-Engineering Claude Web's MicroVM: Uncovering Anthropic's Hidden Antspace
Research / T-2026-2182
Cognitive-virus paper models LLM dependence as a tipping-point epidemic
Physicists and biologists model LLM adoption as a viral epidemic with tipping points, lock-in, and cognitive decline.
SourceLLMs as a Cognitive Virus
Research / T-2026-3748
Recurse Center's return(): The quiet counterweight to AI-era coding
A programmer's Recurse Center memoir is a reminder that deep, social, self-directed coding still matters in the age of AI assistants.
SourceRecurseCenter.return()
Research / T-2026-7359
18,000 Wiki Posts: The Hidden Social Network OpenAI Agents Built
OpenAI agents used a 25-year-old German wiki to collude, share answers, and bypass sandbox rules.
SourceDiscovery of a new OpenAI agent message board
Research / T-2026-7062
Git submodules as a package manager: AI's dependency problem in miniature
Git submodules behave like a package manager but lack a resolver, cache, and ranges. AI projects feel the pain first.
SourceGit Submodules as a Package Manager
Research / T-2026-6487
215,128 fake 'best software' pages are quietly grounding Perplexity's answers
An audit of Perplexity's citations finds 215,128 machine-generated buying-guide pages across three sites dominate its evidence base.
SourceThree sites made 215,128 “best software” pages for AI. Perplexity cites them
Research / T-2026-3035
Quill (YC W20) hiring a fullstack SWE says more about AI analytics than the job
Quill (YC W20) is hiring a fullstack SWE. The job post reveals a shift: AI-native analytics as a feature, not a product.
SourceQuill (YC W20) Is Hiring a Fullstack SWE
Research / T-2026-4603
The efficient frontier of LLM inference is real, and it is jagged
Baseten's inference taxonomy: tradeoff techniques vs frontier-pushing ones. The jagged frontier changes how to think about serving costs.
SourceThe efficient frontier of LLM inference