Another busy day in AI brings significant developments spanning mathematics, security vulnerabilities, coding practices, and the intersection of AI and politics. Here are the top five stories making waves in the AI community today.
1. LLMs Reward Expertise: Domain Knowledge as the Key to Better Prompting
Sean Goedecke’s widely-discussed essay “LLMs Reward Expertise” argues that the most important skill in prompting large language models is — counterintuitively — not prompting technique but domain expertise. Drawing on the famous example of mathematician Terence Tao’s conversation with ChatGPT about the Jacobian Conjecture counterexample, Goedecke demonstrates that subject-matter experts extract dramatically more value from the same models than novices.
Key observations from Tao’s prompting style include: extremely short and direct messages, signaling expertise to push the model into “talking-to-mathematicians” mode, pushing back on wrong responses without directly contradicting, and making independent leaps and suggestions rather than following the model’s proposed direction. The essay argues that the bottleneck in AI-assisted work is increasingly the human, not the model — the information is “in the model” already, but it takes a knowledgeable human to pull it out.
The piece resonated deeply on Hacker News (765 points, 323 comments), with many experienced developers sharing anecdotes about how their codebase familiarity and system design knowledge enabled them to get far better results from LLMs than colleagues without that context. The essay suggests that far from making expertise obsolete, LLMs may actually amplify the value of deep domain knowledge.
2. SQLite Critical CVEs or LLM Slop? Fabricated Vulnerabilities Expose Security Pipeline Flaws
JFrog security researchers published a scathing analysis of recently-reported SQLite vulnerabilities that were initially flagged as critical by NVD and CISA’s ADP. The six CVEs (scored between 7.5 and 9.8) turned out to be entirely fabricated — “LLM slop” generated by AI tools and submitted through MITRE’s public form, which lacks identity verification.
The investigation revealed that the advisories cited non-existent functions, referenced line numbers beyond the end of source files, and described code that had never existed in the targeted versions. One CVE (CVE-2026-51302, initially scored 9.8 Critical) claimed a use-after-free in exprComputeOperands() — a function that didn’t exist in SQLite 3.41.0. Red Hat initially assigned it a 10.0 Critical score before downgrading it to 7.6 High after JFrog’s findings.
The broader issue is systemic: NIST effectively paused deep analysis of vulnerability reports in February 2024 due to a massive surge in submissions, and the pipeline now lacks any requirement for proof-of-concept or bug reproduction. A broader audit of 55 advisories from the same GitHub account found that 54 were completely fabricated. The researchers warn that automated vulnerability triage systems using AI could be particularly vulnerable — an AI agent encountering a fabricated CVE might attempt to generate patches for code that doesn’t exist, wasting time and potentially introducing changes.
3. OpenAI Announces Ten Mathematical Advances with Lean Formalizations
OpenAI published a significant milestone in AI-driven mathematical research, announcing ten advances in mathematics and theoretical computer science, each accompanied by formal proofs verified in the Lean theorem prover. The results, achieved with their latest reasoning model, span a remarkable range of fields:
- High-dimensional sphere packing: Improved asymptotic upper bounds on sphere-packing density, reaching the Cohn–Elkies threshold
- Binary and spherical codes: Exponentially stronger upper bounds for binary codes at every minimum distance
- Non-sofic groups: A construction of a non-sofic group, resolving whether every group admits finite permutation approximations
- Connes’s rigidity conjecture: A counterexample to the conjecture that certain groups are determined by their group von Neumann algebras
- Arithmetic circuit complexity: New lower bounds for computing the permanent, including an n⁴/log n formula lower bound
- Quantum parallel repetition: Exponential parallel repetition for arbitrary finite two-player quantum games
- Closest vector problem: Polynomial-factor hardness of approximation, with consequences for lattice problems
- Ehrhart’s volume conjecture: The sharp maximum volume in every dimension for a convex body whose centroid is its only interior lattice point
- Multicolor Ramsey numbers: A superexponential lower bound, resolving Erdős problem 183
- Extremal number conjectures: Counterexamples to the compactness and degeneracy conjectures, resolving Erdős problems 146 and 180
The accompanying GitHub repository (ten-proofs) includes all Lean formalizations, and the company published a reasoning walkthroughs paper describing how the model reconstructed the proofs. The post generated 514 points and 791 comments on Hacker News, with mathematicians debating the significance of the results and whether the cost figure of ~$2,000 per proof is representative given undisclosed experimental parameters.
4. Preventing Cognitive Debt: Manually Retyping LLM-Generated Code
Ankur Sethi’s provocative essay advocates for an unusual approach to AI-assisted coding: manually retyping every line of LLM-generated code rather than copying and pasting it. The argument is that while AI coding assistants can dramatically accelerate development, they also create “cognitive debt” — a loss of understanding of how one’s codebase works.
Sethi, who describes himself as an experienced developer, uses a system where his coding assistant shows proposed edits in chat, and he manually types them into his editor. He estimates this makes him about 2x faster than working alone, compared to the 10x gains claimed by those who fully delegate to AI. The trade-off, he argues, is worth it: typing the code manually builds a mental model of how it works, helps detect hallucinations and bad design choices, and creates a spatial map of the codebase.
The essay sparked fierce debate (461 points, 372 comments on HN). Some commenters argued that if the workflow involves thinking hard, letting the LLM write, reviewing, and then retyping, the efficiency gains are questionable. Others echoed Sethi’s concern about the industry taking on massive cognitive debt, warning that large parts of digital infrastructure could soon be understood by no one. The piece draws a parallel to the traditional programming advice of never copying and pasting code without understanding it — with LLMs, that advice may be more relevant than ever.
5. OpenAI’s Super PAC Linked to AI-Generated News Site Targeting Industry Critics
An investigation by Model Republic’s Tyler Johnston reveals that OpenAI’s $125 million super PAC, Leading The Future, appears to be funding an AI-generated news site called Acutus (The Wire by Acutus) that publishes articles attacking AI industry critics. The investigation began when Encode’s vice president and general counsel, Nathan Calvin, received a suspicious interview request from a “Michael Chen” — a reporter who turned out to be an AI agent.
The site, which launched in December 2025, has published 94 articles. Analysis with the Pangram AI detector found 69% were fully AI-generated and 28% partially AI-generated — only 3 articles were classified as human-authored. The site’s JavaScript code reveals an editorial dashboard with fields for “AI Background Context” and “Question Prompts,” a “Generate Story Draft” button, and an automated review system that scores articles on AP style compliance, quote accuracy, and source verification.
The investigation connects Acutus to Patrick Hynes, president of Novus Public Affairs, a GOP PR firm whose client list includes Targeted Victory — the firm at the center of OpenAI’s political apparatus. Hynes’ firm also represents PhRMA, the pharmaceutical lobby group whose interests align with Acutus’ coverage. The site’s articles attack AI safety advocates, criticize both blue and red state AI regulation, and generally align with the anti-regulation lobbying positions of Leading The Future.
The story raises serious questions about AI companies using their own technology to generate political propaganda under the guise of independent journalism — a practice that OpenAI’s own usage policies explicitly prohibit. The revelation comes alongside other recent reports of OpenAI’s astroturfing efforts, including a children’s safety coalition and a massive grassroots supporter list generated through paid advertising.
Closing Thoughts
Today’s stories capture the full spectrum of AI’s impact: from genuine scientific breakthroughs in mathematics to systemic vulnerabilities in security infrastructure, from debates about how best to work with AI tools to hard questions about the ethics of AI-generated political content. As AI continues to reshape every domain it touches, the tension between its potential for discovery and its potential for misuse remains the defining story of our era.
