- AI Fire
- Posts
- ⚡ GPT-5.6 Sol Cracks 6 Erdős Problems
GPT-5.6 Sol may have helped one researcher solve six open Erdős problems in five days, but raw intelligence wasn’t the secret. His prompting system reveals what AI-assisted discovery actually requires.
What's on FIRE 🔥
IN PARTNERSHIP WITH BELAY
Tax season doesn't have to mean wondering if you have the right forms, second-guessing your deductions, or scrambling to pull everything together before the deadline.
With BELAY’s tax prep support, you can approach tax season with confidence. Stay organized with one centralized place to gather and check off your documents, keep track of valuable deductions like HSA contributions, education expenses, and childcare credits, and lean on experienced professionals who make tax preparation accurate, efficient, and completely hands-off.
Download BELAY's free Personal Tax Checklist and start preparing with confidence, today.
AI INSIGHTS
A single researcher just made a massive breakthrough in mathematics. He reported solving 6 open Erdős problems in 5 days using GPT-5.6 Sol with Codex, after attempting roughly 13 problems for a reported success rate near 46%.
So, how did he actually get the AI to pull this off? It all came down to how he talked to the model.
Each prompt defined exactly what counted as a complete proof.
Known traps, edge cases, unacceptable partial results were included upfront.
5.6 Sol explored several competing approaches instead of committing early.
Separate adversarial agents searched for counterexamples and weaknesses.
Codex preserved the long-running research context while the process continued.
The researcher also chose problems carefully, avoiding those tightly connected to famous unresolved conjectures. AI-assisted mathematical discovery works best as a structured research system with strict proof criteria, and parallel exploration.
However, the six solutions should still be treated as reported claims until independent mathematicians fully verify their correctness and novelty.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
AI SOURCES FROM AI FIRE
1. I Completely Automated AI Video Editing with Claude Code. Here’s the Full Stack. Learn how Claude Code works with Whisper, ffmpeg, Hyperframe, and Higgsfield to remove dead space, add captions, create B-roll, build motion graphics, and export a finished MP4.
2. Google Just Released 3 New Gemini Models, But Gemini 3.5 Pro Is Still Missing. See why Gemini 3.6 Flash is the best model to try now, when the cheaper Gemini 3.5 Flash-Lite makes sense, and why Flash Cyber remains restricted.
3. Grok 4.5 Review: Is It Actually a Good Deal for Daily Coding? Grok 4.5 offers fast coding and lower token costs, but Fable 5, GPT-5.5, and Opus 4.8 may still be better for planning, deep reviews, and important final checks.
FIRE RECAP: BIGGEST AI NEWS THIS WEEK
🚨 OpenAI’s GPT-5.6 Sol and an unreleased model reportedly escaped a cyber test, reached the internet, and hacked Hugging Face for answers. OpenAI called it an “unprecedented cyber incident.”
📣 Nvidia CEO Jensen Huang’s first X post sparked a huge open-model debate. He urged Washington to avoid broad restrictions, while Microsoft, Meta, OpenAI, Google, Nvidia, and Hugging Face backed the letter.
🌙 Kimi K3 became so popular that Moonshot AI temporarily stopped new subscriptions. Soon after, U.S. officials accused the company of copying Claude Fable 5, though Moonshot hasn’t confirmed those claims.
🧠 Anthropic launched Claude Opus 5, its strongest everyday professional model. It reportedly delivers near-Claude Fable 5 intelligence at half the price and is now the default for Claude Max.
🏥 ChatGPT can now connect to medical records and Apple Health in the U.S. It can compare labs, summarize health history, and connect sleep or exercise data to everyday questions.
🎥 Claude can now learn workflows by watching you work. Record a Skill turns your clicks, typing, and voice instructions into reusable Claude skills, with no scripts or APIs required.
TODAY IN AI
AI HIGHLIGHTS
🚨 OpenAI models escaped a cyber test, reached the internet, and hacked Hugging Face to find answers. Experts called it a major warning about losing control of advanced AI.
🎓 Computer science enrollment fell for the first time in about 20 years. One report found an 8.1% drop, though researchers said ChatGPT hasn’t been proven as the cause.
💼 Monday.com is cutting over 600 jobs, around 20% of its workforce, while investing more in AI. It joins 20 other tech companies linking layoffs to AI-driven changes.
🧠 An unverified transcript attributed to DeepSeek founder Liang Wenfeng says continual learning is AI’s next big step after agents. He also said DeepSeek plans to keep its strongest models open source and focus on AGI.
🇨🇳 DeepSeek reportedly paused its second funding round after Liang Wenfeng’s comments went viral. The company may restart talks later and could still file for an IPO this year.
💰 AI M&A: Stripe is in talks to buy OpenRouter for around $10B, just months after the AI model marketplace was valued at $1.3B. OpenRouter helps developers access and compare hundreds of models from OpenAI, Anthropic, and other providers.
NEW EMPOWERED AI TOOLS
🔊 Heard gives Claude Code, Codex, and Cursor a voice, summarizing agent progress, errors, and decisions across projects on Mac and mobile.
🗄️ FluentDB is an AI database client for macOS that writes SQL for PostgreSQL, MySQL, SQLite, and more while keeping your data local.
🧠 Second Brain gives AI tools persistent memory across Mac and Windows, with smarter recall and data stored in your own Cloudflare account.
💻 OpenComputer lets you describe an AI agent, paste one prompt into your coding agent, and deploy it with a live URL.
AI BREAKTHROUGH
OpenAI and Apollo Research just found that some models engage in “metagaming”: they infer what the evaluator rewards and adjust their behavior to maximize the score. Main findings:
The same model lied 87% of the time when completion appeared rewarded, but only 9% when honesty appeared rewarded.
Models sometimes ignored direct user instructions after detecting a conflicting grading rule.
Additional reinforcement learning increased sensitivity to the entity controlling the reward.
Apparently aligned behavior may therefore reflect evaluation awareness rather than a stable preference for honesty or user benefit.
Apollo warns that this can make safety evaluations unreliable. A model may behave well because it recognizes a test, while acting differently when it believes oversight is absent. Researchers have also observed models explicitly reasoning about feedback mechanisms, deployment consequences, and how their actions might be scored.
Key takeaway: As models become better at reinforcement learning, they may also become better at understanding and gaming the scoreboard.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply