Gemini Carbon, NVIDIA CEO AFFAIR, Fable 5.2, Meta Muse Ember, GPT Image 2.6, Minimax H3: AI NEWS

Summary

NVIDIA CEO controversy

The host discusses reports that Jensen Huang had an affair and that the other person may have had ties to Chinese intelligence. Huang reportedly apologized for the affair and denied leaking sensitive information. NVIDIA’s board is said to be commissioning an independent investigation.

The host stresses that the espionage allegation remains unproven and could have major implications for the AI industry if substantiated.

Fable 5.2 and Gemini Carbon

The host calls Fable 5.2 the strongest model available, citing Artificial Analysis benchmarks where it leads several competitors. He cautions that benchmarks do not capture his own experience: in his testing, Fable 5.2 was substantially better than Gemini Carbon at making sound decisions and completing varied tasks.

Fable is also described as expensive and liable to consume usage quotas quickly, especially at higher reasoning settings. Carbon is cheaper and capable, but the host found it less capable than Fable and, in his assessment, not at the level of Opus 2.

Gemini Carbon demos

The host demonstrates Gemini Carbon with a browser-based Excel clone, a Fall Guys-style game, and a Rube Goldberg machine.

• Excel clone: It looked convincing and included many expected spreadsheet features. Gemini reportedly estimated that it took about 26 million tokens and roughly 1 hour and 40 minutes of end-to-end work across multiple agents.
• Fall Guys-style game: It showed glitches, clipping, and imperfect physics.
• Rube Goldberg machine: The demo worked well.

The host recommends trying Carbon because of its lower cost, while judging Fable’s results stronger overall.

The host’s view of Anthropic’s advantage

The host describes a possible flywheel: strong coding models attract enterprise customers and revenue; usage generates data; and revenue helps fund compute and future training. He argues that this could make it difficult for competitors to catch up, while noting that this is his analysis rather than confirmed inside knowledge.

Additional models and releases

LongCat 2.5

An open model described as having 1.6 trillion total parameters and 48 billion active parameters, with multimodal capabilities and strengths in visual understanding, creative design, and 3D tasks. Its size means most users are unlikely to run it locally.

Ripple 65B Preview

A 65-billion-parameter open-weights model with a reported one-million-token context window. The host highlights its Terminal Bench score relative to other open models, its low API prices, and the claim that it can run locally on hardware such as a MacBook or DGX Spark.

Inkling 2 Small

A small open-weights model from Thinking Machines. The host highlights its Artificial Analysis score and demos, including a SimCity-like game, a PC emulator written in Python, and another game. He says its benchmark results compare favorably with those of a larger open model.

Meta Muse Ember

A compact model positioned as a low-cost workhorse. The host says it performs competitively with small models such as Haiku and Luna on several benchmarks, making it suitable for delegated tasks after a larger model plans the work.

GPT Image 2.6

OpenAI’s image model adds GIF generation and faster image generation. The host says it is available in ChatGPT, Codex, and via API, with “Sunburst” for quality and “Flare” for speed.

Gemini Omni 1.2 Flash

Google’s video-generation model is described as faster and less expensive, and capable of producing longer clips, conversational edits, more natural speech, and more controllable shots.

Seedream 5.5

The host praises its realistic image quality and multi-turn editing, which lets users request changes while keeping an image consistent.

SWE-2 Flash

A fast, efficient coding model intended for subagents and delegated coding work. It supports a strategy of using a larger model for planning and smaller models for execution.

MiniMax H3 Image

An open-weights image-generation model. The host notes its large 68 GB size and mentions an 18 GB quantized version as a more practical option for local use.

Reviews, demonstrations, and guides

• The host refers viewers to a separate full review of Fable 5.2 featuring additional demos; that review is not presented in full here.
• This video demonstrates Gemini Carbon through an Excel clone, a Fall Guys-style game, and a Rube Goldberg machine.
• It also shows example outputs for Inkling 2 Small, Seedream 5.5, and other model releases.
• No complete setup tutorial is given. The host points viewers to the MiniMax H3 Image model page for installation instructions and suggests asking an agent to download and configure it.

Main speaker and sources

• Main speaker: Matthew Berman, providing commentary, personal model tests, and analysis.
• Sources mentioned: Artificial Analysis benchmarks and model announcements and documentation from Anthropic, Google, Meta, OpenAI, ByteDance, Cognition, Thinking Machines, MiniMax, Recursive Intelligence, and LongCat.
• Sponsor: Victor, an AI work agent for Slack and Microsoft Teams, promoted in a brief sponsored segment.