|
| |
| |
Good morning, AI enthusiast. |
Mark Zuckerberg just published a 6,500-word case against letting a handful of companies control superintelligence - and backed it up with something concrete: Muse Glimmer, a 30-billion-parameter model anyone can download and run on their own laptop. |
It's Meta's clearest bet yet that owning your AI beats renting someone else's, at a moment when Anthropic is pushing the frontier of what models can do and OpenAI is locking down its most capable ones. If the smartest models get harder to access even as the useful ones get easier to own, which strategy actually wins? |
Today in AI Brief: |
Zuckerberg bets big on open-weight AI Claude pushes a Riemann bound past 67% OpenAI arms defenders with a hacking model |
|
|
|
| |
|
| |
| |
Zuckerberg Bets on Open AI With Muse Glimmer |
In Brief: Meta released Muse Glimmer, a 30-billion-parameter open-weight agentic model, alongside a 6,500-word essay from Mark Zuckerberg arguing superintelligence should stay broadly accessible rather than concentrated among a handful of labs. |
The Details: |
Glimmer runs on one consumer GPU - a quantized version needs under 20GB of memory and works on a MacBook M4-Max or an RTX 5090, no data center required. It outperforms Gemma4-31B and Qwen3.6-27B on agentic, coding, and reasoning benchmarks, and ships with tool use, multi-step reasoning, and failure recovery built in. Zuckerberg argues concentrated AI power is inherently risky, writing there's "no such thing as a singular benevolent superintelligence," and commits Meta to keep releasing open-weight models. |
Take Away: |
Meta is betting that owning your AI beats renting someone else's - download Glimmer, and it runs offline with no vendor in the loop. That's a direct challenge to the closed-model strategies at OpenAI and Google as the open-versus-closed fight defines the next phase of the AI race. |
|
|
|
| |
|
| |
| |
Claude Cracks a Piece of the Riemann Hypothesis |
In Brief: An unreleased research version of Claude raised a century-old bound tied to the Riemann hypothesis from 41.6% to 67.2% - the first time the proven share of zeta zeros on the critical line has cleared half. |
The Details: |
Claude got there by running two Claude Code sessions that spun up roughly 60 subagents, wrote hundreds of Python scripts, and burned through 31 million output tokens chasing the result. The first attempt produced 650 dead-end ideas before a second pass found the winning approach, which two Anthropic mathematicians then verified and formalized as a machine-checkable proof. Outside mathematicians Brian Conrey and Dan Goldston reviewed the work independently, and Claude cross-checked 54 arXiv papers to confirm the result was actually new. |
Take Away: |
Claude didn't solve the Riemann hypothesis, and Anthropic doesn't expect its methods to get there. But turning weeks of expert human effort into two automated sessions shows AI can now push the edge of open math problems, not just summarize them. |
|
|
|
| |
|
| |
| |
OpenAI Arms Cyber Defenders With GPT-5.6-Cyber |
In Brief: OpenAI expanded its Daybreak program into two access tiers and introduced GPT-5.6-Cyber, a purpose-built model for vetted security researchers hunting exploits and validating defenses. |
The Details: |
GPT-5.6-Cyber completed 95% of advanced cyberattack prompts in internal testing - things like exploit-chain development and privilege escalation - versus just 1.5% for the standard safeguarded model. Daybreak Blue gives partners access to GPT-5.6 Sol without its default cyber guardrails, while Daybreak Red unlocks the full Cyber model for exploit validation and vulnerability research. OpenAI widened access to major security firms including Accenture, IBM, and CrowdStrike, and will require all individual Daybreak accounts to use hardware security keys starting September 1. |
Take Away: |
Handing defenders a model that can do what attackers can do is a bet that better tools for the good guys outrun the bad ones. With OpenAI already treating its next model as "critical" risk for cyber capabilities, this is the industry's clearest signal yet that AI-versus-AI cybersecurity is now the default. |
|
|
|
| |
|
| |
| |
Everything else in AI |
Dyna Robotics unveiled Dyna-2, a robot foundation model trained on over 1 million hours of human video that hits an 87% zero-shot task success rate at new customer sites, up from 46% for its predecessor. |
Discovered Materials is using Anthropic-powered AI agents alongside physics simulations to generate thousands of candidate chip-cooling materials a day, after raising a $9M seed round from Lightspeed India and Peak XV. |
Spotify launched Xirp, a vendor-neutral platform for running dozens of AI coding-agent sessions at once, already adopted across more than 36,000 sessions by its own engineers. |
New York and Texas joined a wave of local pushback that has pushed the number of U.S. jurisdictions restricting new data centers past 500. |
|
|
|
| |
|
|
|
