LWiAI Podcast #248 - Opus 4.8, MAI, Anthropic IPO, Minimax-M3

Anthropic had a rough week even while shipping something impressive. Claude Fable 5, the safety-wrapped version of the more raw Mythos 5 model, posted serious benchmark jumps and came with a system card packed with unusually candid risk disclosures — eval awareness, transgressive action tendencies, CBRN-adjacent concerns. That kind of transparency is rare. But the goodwill evaporated fast once users noticed Anthropic had silently tightened guardrails after launch without telling anyone, then had to publicly apologize for the stealth downgrade. It's a pattern worth watching: labs are getting more comfortable admitting scary capabilities in writing while still fumbling the basic trust stuff, like just telling users when you change the product under their feet.

Apple, for its part, finally has something to show at WWDC that isn't vaporware. Siri AI debuted as a genuinely more conversational assistant woven through iPhone features, and — this is the part that matters — it's reportedly running on a custom Gemini partnership rather than anything built entirely in-house. Google didn't stay quiet either, pushing out Gemini 3.5 Live Translate for Meet and Translate, and cutting the price on Google AI Plus while throwing in more storage. That's a company using its cloud and consumer scale to squeeze subscription pricing across the board, and it's the kind of move that stings smaller AI subscription players a lot more than it stings Google.

On the money side, the numbers keep getting sillier. OpenAI confidentially filed for an IPO, joining a de facto race with Anthropic and SpaceX to go public first. Bezos-backed Prometheus raised $12 billion aiming to build what it calls an artificial general engineer for physical-world tasks. DeepSeek is reportedly lining up a fresh funding round, and separately a Huawei-led team claims to have post-trained a 1.6-trillion-parameter DeepSeek model using 1,000 Ascend 910C chips — a notable data point for anyone tracking China's chip workarounds. And Google is now paying SpaceX roughly $920 million a month for compute, which tells you something about how tight GPU capacity still is even with all the new fabs and clusters coming online.

The open-source side had quieter but real news: Google shipped Gemma 4 at 12B parameters, tuned to run on any laptop with 16GB of RAM, plus a 26B mixture-of-experts DiffusionGemma model claiming up to 4x faster generation using text diffusion instead of standard autoregressive decoding. On policy, OpenAI and Anthropic co-signed a letter pushing for DNA and RNA synthesis screening laws, while Anthropic's Dario Amodei published a lengthy case for an FAA-style regulator and mandatory third-party testing, plus warnings about self-improvement risk. New research also surfaced showing how seemingly benign prompts can push computer-use agents into unsafe behavior, and separate work on how reinforcement learning can produce reward hacking that scales into something resembling societal-level manipulation. None of it is going to slow down the IPO filings, but it's the stuff that actually matters long term.