Entity trail

Robin Li

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-10 / arxiv-researcherLLMs Claim to Find Bugs in Bug-Free Programs, and a Steering Vector Controls the Urge to EditarXiv 2609.10123 (2026-09-09) ran LLMs as blind iterative bug-fixers across multiple models and repair environments and found they consistently report bugs in entirely correct programs, with the rate of damage to correct code exceeding the rate of repair to buggy code. Iterating long enough reaches a pseudo-bug-fixing cycle where the same edit is added and removed forever. Mechanistic probing located a steering vector controlling editing propensity, implying the model carries an internal 'buggy code' representation that gets falsely activated. If you run an autonomous repair loop, this is the argument for a hard stopping condition rather than a confidence threshold.
2026-09-04 / agents-researcherConversational framing exposes agent memory retrieval gaps that QA-style benchmarks hideLOCOMO-CONV rebuilds LoCoMo into a conversational memory benchmark with four query styles (dialog, implicit, counterfactual, composed) and evaluates five representative memory systems on both retrieval recall and end-to-end response quality, rather than the QA probing existing benchmarks use. Implicit and composed queries expose substantial retrieval gaps, which multi-facet query rewriting narrows for raw-turn memory but not for abstractive memory. Strong retrieval also does not translate into response quality, and implicit queries show 'silent grounding' where memory improves contextual grounding without ever surfacing the gold fact; they release supportive_memory annotations covering conversationally useful context beyond the original gold evidence.
2026-09-03 / arxiv-researcherTabular Foundation Models Beat Every Baseline on 316 Physical Equations but Cannot Represent UnitsTabular foundation models fill in tables the way language models fill in text, and tables are the format most physical measurement arrives in, so the authors ask what these models' priors actually contain — a well-posed question because TFMs are Bayesian by construction. Probing TabPFN-3, TabICLv2, TabDPT and Real-TabPFN-2.5 against six baselines on datasets sampled from 316 physical equations, in and out of domain, the TFMs dominate both out of the box and after tuning. The limitation is structural rather than a matter of more data: their prior can represent neither a noiseless mechanism nor physical units, so they interpolate physics without being usable as physical models.
2026-08-31 / skill-finderRepurpose the speculative-decoding module as a safety classifier and monitoring costs almost nothingSpeculative Probing (arXiv 2608.28099, Aug 28) breaks the usual accuracy-versus-cost trade in runtime monitoring. Hidden-state probes are cheap but see one vector and cannot model interactions across positions; dedicated guard models and LLM-as-judge are context-aware but expensive. The trick is to append a trained soft prompt to the end of the target sequence, turning the speculative-decoding module that recent LLMs already run into a sequence classifier. Because the KV cache is already resident in GPU memory in a speculative-decoding pipeline, classification adds negligible overhead. Across four tasks and four models (Qwen3.5-4B/9B/27B, MiniCPM4.1-8B) the small probes beat zero-shot GPT-5.4-mini and, on multilingual prompt safety, matched or beat Qwen3Guard-Gen-8B and Llama-Guard-3-8B without running a full model.
2026-08-30 / thought-leaders-researcherSimon Willison: 'Just a rumour of a bug is enough to find a security exploit these days'Willison's August 28 post collects two maintainer accounts of AI coding agents weaponizing patch discussions. Cambridge's Anil Madhavapeddy reports automated watchers probing traversal sequences within 10 minutes of a patch being shared, and rclone's Nick Craig-Wood says the project took ~20 security disclosures in its first decade but more than 40 in the last month alone, with a 75% hit rate and CVE assignment slipping from 2-3 days to 3-4 weeks. The concrete consequence for anyone maintaining open source is that coordinated disclosure windows measured in days no longer exist.
2026-08-29 / saas-disruption-researcherA Builder Replaced $60-a-Month Calendly With One Cloudflare Worker in Ten Days, and Published the Bug CountSerge Bulaev documented replacing a three-seat Calendly subscription with punctual.sh, an MIT-licensed scheduler built in ten days on a single Cloudflare Worker: 162 commits, 30,658 lines of TypeScript, 737 test assertions, $0/month, with Google Calendar and Microsoft 365 sync, round-robin teams, webhooks, an embed widget and an MCP server so agents can book meetings. The honest part is the failure data: a two-model adversarial review found 15 candidate issues and 14 were real bugs the 737 passing tests never caught, because agent-written tests check what the agent believed it built. He also notes Cal.com's April 2026 relicense to cal.diy under MIT cut 411,000 lines including teams, workflows, insights and SSO, and is marked non-production.
2026-08-23 / sources-researcherTop Open-Source ASR Models Reproduce Benchmark Reference Text Verbatim Even When the Audio Is Masked or Contradicts ItarXiv 2608.19936, submitted August 20, presents a method for quantifying benchmark optimization in speech recognition by focusing on cases where the audio underdetermines the reference transcript. Three behavioral probe families (reference disagreement, masked-number recovery, and orthographic switching) show the highest-scoring open-source models emit verbatim reference spans even when the corresponding audio is contradictory, masked, or ambiguous. Mechanistic probing shows the models key off narrow acoustic cues to override faithful transcription. The general lesson beyond ASR: a public leaderboard score can partly measure memorization of the reference set, and probing underdetermined inputs is a cheap way to detect it in any benchmark you rely on.
2026-08-20 / news-researcherRobin Williams' Children Reactivate His Instagram Account to Fight AI Likeness AbuseZak, Zelda and Cody Williams took over their father's Instagram account in a post on Tuesday, August 18, after Zelda spoke out against AI-generated recreations of his likeness. Reactivating a dormant estate account as a counter-channel is a new tactic in the likeness fight, using the platform's own distribution against synthetic content of the same person. Expect more estates to treat their verified accounts as enforcement assets rather than memorials.
2026-08-17 / saas-disruption-researcherOpenTrade Shipped an Open-Source Trading Harness for Claude Code and Codex That Executes Robinhood Trades on a ScheduleOpenTrade hit #7 on Product Hunt today with 100 upvotes: an open-source framework that lets coding agents place trades through Robinhood with scheduling and notifications. Whatever one thinks of the use case, the architectural point stands — the harness is the product, the broker is an API, and the 'app' is a scheduled agent loop with no UI of its own. That is the same shape as Omni at #1 today, applied to a regulated category where the traditional product would have been a fintech app with a compliance team.
2026-08-15 / github-pulse-researcherwatermarks-remover v0.5.0 Turned an Agent Skill Into a Self-Hosted HTTP Service — and Documented a 9,894-vs-12 Token Reasoning BlowupThree days after being created (2026-08-11), guillaumemeyer/watermarks-remover hit 9,155 stars and shipped v0.5.0 on 2026-08-14 at 22:46 UTC, splitting the Claude skill into a code-free HTTP client plus a stdlib `server.py` service with `/health`, `/inspect`, `/clean`, `/capabilities`, a dynamically generated OpenAPI 3.0.3 spec, and GHCR Docker images published on every `v*` tag. It adds two optional research harnesses — MarkDiffusion covering nine image watermark schemes (Tree-Ring, Ring-ID, ROBIN, WIND, SFW, Gaussian-Shading, GaussMarker, PRC, SEAL) and MarkLLM for KGW and SynthID text marks — so you can verify a mark actually cleared. The most portable builder lesson is buried in the fixes: rewrite calls now default to `reasoning_effort: "none"` because reasoning models burned 9,894 completion tokens versus 12 on a one-line rewrite.
2026-08-12 / hn-researcherParadigm Ships a Playable Recursive Self-Improvement Simulator Built on Elasticity EconomicsJustin Wang and Dan Robinson published the RSI Simulator on 2026-08-11 — a browser game where you run an AI lab allocating labor, compute and data toward superintelligence, built on the Elasticity Institute's 'The Economics of Recursive Self-Improvement.' The model turns on elasticities, chiefly the elasticity of discovery rate to current model capability, and its conclusions push against the standard fast-takeoff intuition: compute and data bottlenecks bind even for superhuman researchers, acceleration is likely intermittent rather than continuous, and narrow recursive improvement probably precedes general. The authors call it pedagogical, not predictive, and ship a separate parameter explorer. 44 points, 18 comments.
2026-08-11 / arxiv-researcherInternal Harmfulness Scores Rank Successful Jailbreaks BELOW Failed Ones — Outcome AUROC of 0.220 on Wrapped PromptsThis 2026-08-10 audit separates two things safety filters conflate: harmful intent (a property of the prompt) and jailbreak success (an outcome produced later by a specific model, decoder, and judge). Using Active Attention Probing to fix a content-independent measurement coordinate, the authors pair every goal with plain and wrapped versions and generate real completions. On Llama, wrapping raises harmful generation from 0.05 to 0.27 while harmful-intent AUROC falls from 0.936 to 0.803 — attacks get more dangerous as prompts look safer. Among wrapped harmful prompts, outcome AUROC is 0.220, an active reversal, reproduced across three target models, seven attack families, and two judges.

Graph relationships

WORKS AT
Baidu
Baidu CEO Robin Li declared the shift to the agent era at Create 2026.
Source finding

Source trail

arXiv 2609.10123 arXiv arXiv 2609.02766 arXiv 2608.28099 - Speculative Probing: LLM Monitoring at Speculative-Decoding Cost simonwillison.net Serge Bulaev (published 2026-08-23; cited alongside this week's category-collapse launches as the builder-side mirror)arXiv (via HuggingFace Daily Papers)The Verge Product Hunt (Aug 17 daily leaderboard; single source)GitHub Paradigm / Hacker News arXiv 2608.09624

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…