Tech • AI • Robotics • Game

ENFR

GLM 5.5 Is Coming Soon! Claude Cracks 167-Year-Old Math Problem?! Google Loop & More! AI News

9.4/10
AIWorldofAINovember 21, 2026 at 08:12 AM32:27
Audio player

⏱ ≈ 3 min read · 32-min video — 29 min saved

TL;DR

AI companies pushed ahead on multiple fronts, from an unverified claim of a Riemann Hypothesis proof aided by Claude to new model launches, autonomous coding agents, and broader use of AI in scientific research.

KEY POINTS

Claimed Riemann breakthrough

Anthropic mathematician Levent Alpoge has claimed a proof of the Riemann Hypothesis, a 167-year-old unsolved problem central to number theory and the distribution of prime numbers. Reports say the work was developed with significant help from Claude over roughly six weeks, and that the same approach may also extend to the Generalized Riemann Hypothesis.

Formal verification raises the stakes

The proof has reportedly been expressed in Lean, a formal proof language that allows computers to mechanically verify each logical step. Circulating descriptions put the project at about 17 million lines of Lean code and more than 250,000 pages of human-readable mathematics, making it an unusually large AI-assisted mathematical construction. The claim remains subject to independent scrutiny, including whether the formalized result exactly matches the headline theorem.

Why the Riemann claim matters

The Riemann Hypothesis, proposed in 1859, concerns the non-trivial zeros of the Riemann zeta function and has deep implications for estimates of how prime numbers are distributed. A valid proof would settle one of the Clay Mathematics Institute’s Millennium Prize Problems and strengthen a wide range of results in number theory, algorithms, and theoretical computer science that currently rely on the hypothesis being true.

Possible new Chinese flagship model

Zhipu AI appears to be preparing a new model before December, fueling speculation that GLM 5.5 could be close. The timing stands out because it has been about 98 days since GLM 5.3, longer than the company’s roughly 61-day average flagship cadence this year.

Mystery around Gestalt

Speculation intensified after a stealth model called Gestalt topped a design benchmark with an apparent 1900 Elo and 90 percent success rate. That would place it well ahead of human designers at around 1700 Elo and ahead of Fable 5.5 at 1713, though benchmark leadership alone does not establish broader reasoning or coding strength.

Google explores autonomous software work

Google is reportedly developing Sloop, an internally tested goal-driven agent tied to Jules v2. Instead of completing one-off tasks, it is designed to connect to a GitHub repository, define long-term objectives such as security or engagement improvements, break them into smaller tasks, deploy changes with Gemini, and then monitor metrics to evaluate whether the product actually improved.

Inkling 2 Mini targets efficient local AI

Thinking Machines released Inkling 2 Mini, a 26 billion-parameter open-weight model that scored 20.1 on the Artificial Analysis Intelligence Index. Its looping-transformer design reuses middle layers rather than simply expanding parameter count, and only 3.4 billion parameters are active at a time, making it notable for efficient local inference.

Open licensing and local deployment

Inkling 2 Mini is released under Apache 2.0, allowing modification, fine-tuning, and commercial use. It can reportedly run locally on a Mac with 32 GB of RAM, highlighting a broader shift from pure scale toward more compute-efficient model architectures.

Product updates across major labs

Claude’s mobile app is getting a redesign focused on code projects and artifacts, while OpenAI expanded ChatGPT Spaces to Plus users for collaborative presentations, documents, and interactive pages. Google also added new Gemini features including AI-generated GIFs, voice-based editing through the Gemini API, and multitasking improvements in Antigravity.

AI expands into science and accessibility

OpenAI said its Rosalind effort is being used to support research on neurodegenerative diseases including Alzheimer’s and Parkinson’s, alongside a MedChem benchmark for medical chemistry reasoning. Separately, Neuralink demonstrated a brain-computer interface helping a person with paralysis play GTA 6, underscoring how AI-adjacent technologies are beginning to affect both research and assistive computing.

CONCLUSION

The biggest development is the still-unverified claim that AI-assisted work may have cracked one of mathematics’ hardest problems. Even if that result does not hold, the broader pattern is clear: AI is moving beyond chat and coding into discovery, autonomous software work, and real-world scientific and accessibility applications.

Ask a question

More from AI