AI is no longer just answering questions. It's doing the work.
Claude failed to prove the Riemann Hypothesis (obviously) but stumbled onto the biggest single jump in 160 years of math history. OpenAI's GPT-5.6-Cyber found zero-days in Chrome before most humans even knew to look. Both unreleased, both already outperforming the field.
The pattern is clear: the most interesting AI results are happening before the models ship.
|
|
In Partnership with
|
·
|
|
|
|
|
|
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
|
|
Top News
|
|
|
|
Anthropic's unreleased Claude model pushes Riemann hypothesis progress from 41.6% to 67.2%
|
|
16,928 Likes
|
|
Someone told Claude to take a real stab at one of math's hardest unsolved problems. Claude failed. But what it found on the way down is interesting. The Riemann Hypothesis is a 160-year-old math problem about where a special function hits zero. Think of it like a treasure map where all the X marks should land on one specific line. Nobody has proven they all do. The best humans could do was prove that at least 41.6% of those X marks land where expected. Claude just pushed that to 67.2%, the biggest single jump in history on this problem. It did this by treating the zeros as one unified space instead of analyzing them separately. - Used 60 AI subagents running in parallel inside Claude Code
- Tried 650 failed ideas before finding the approach that worked
- Ran 2,400 shell commands over a day and a half
- Proof verified in Lean 4 and reviewed by two external math experts
The proof code and process logs are all public. The model is unreleased, but the agent architecture is yours to study today.
|
|
|
|
Presented by Agent Field AI
|
|
#1 Open-Source Code Review, ~10x Cheaper
|
|
PR-AF just posted the #1 open-source score on Martian's Code-Review-Bench: 0.706 across 42 tools, ~3x more valid findings than leading commercial tools.
Built for deep code review, not shallow diff summaries: it plans a strategy per PR, runs reviewer agents in parallel, and verifies every finding against your source.
Use cheaper models for regular PRs, GLM-5.2 for CI gates, or Opus-class for highest-stakes reviews, at ~10x cheaper per review. Licensed Apache 2.0, self-hosted, and one-step GitHub Action.
|
|
|
|
Top News
|
|
|
|
OpenAI ships GPT-5.6-Cyber, a model that found zero-days in Chrome's V8 engine
|
|
9,319 Likes
|
|
OpenAI just split its Daybreak cybersecurity program into two tiers, and the difference matters a lot depending on what you're building. Daybreak Blue is the starting point. You get GPT-5.6 Sol with security guardrails loosened just enough for defensive work. Think of it as the standard model, but it won't refuse you when you ask it to analyze malware or review vulnerable code. Daybreak Red is the serious tier. It unlocks GPT-5.6-Cyber, a model specifically trained to do the risky stuff: finding zero-day vulnerabilities (unknown security holes), building exploit chains, and validating attacks. It completed 95% of advanced security tasks in testing, versus just 1.5% for the regular GPT-5.6. - Blue: vulnerability discovery, malware analysis, code review, incident response
- Red: exploit development, zero-day research, security testing
- Real proof: it found two unknown bugs in Chrome's V8 JavaScript engine
Red access requires approval. You apply, get vetted, then get monitored. Not open to everyone yet.
|
|
|
|
Presented by Datadog
|
|
Lead AI From Pilot to Proven in 90 Days
|
|
AI has moved from pilot to production, and engineering leaders are expected to show results. This playbook gives you a practical framework of four priority plays, three enabling capabilities, and a 90-day checklist to build reliable, evidence-driven AI systems.
|
|
|
|
Top News
|
|
|
|
NVIDIA releases open 30B model that runs 4x faster than similar-sized models
|
|
3,469 Likes
|
|
NVIDIA just dropped Nemotron 3.5 Lightning, and the story here is speed without the usual tradeoff. Most AI agents waste compute. They use a massive, slow model for every single step, even boring ones like calling a tool or checking a result. Lightning fixes that. Here is the trick: it has 30B total parameters but only activates 3B at a time. Think of it like a team of specialists where only the right people show up for each task, instead of the whole company attending every meeting. That is the MoE (Mixture-of-Experts) design. The numbers back it up: - 4x faster output than similar-sized models
- 86% accuracy on PinchBench, completing 10,000 tasks 35% faster than Qwen3 35B
- 1M token context window for long agent sessions
- Runs on a single H100 or DGX Spark
You can fine-tune it on your own data using NVIDIA NeMo for domains like cybersecurity, coding, or legal work. Weights are open on Hugging Face, deployable via vLLM or SGLang today.
|
|
|
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 300,000 developers stay informed and ahead. We're passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.
Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We're based in the United States. By using our site, you agree to be governed by U.S. laws.
|
|
|
Looking to promote your company, product, service, or event to 300,000+ AI developers?
|
|
|
|
|