LWiAI Podcast #243 - GPT 5.5, DeepSeek V4, AI safety sabotage

Another week, another pile of model drops, and this one's got a genuinely weird detail buried in it: GPT-5.5's system card reportedly includes a warning about "goblins" in the system prompt. Nobody on the podcast fully explains what that means, which is sort of perfect for where AI safety documentation is right now — dense technical papers on chain-of-thought monitorability sitting right next to language that sounds like it escaped from a fantasy novel. Coding improvements are the headline feature, but so is a price hike over GPT-5.4, so OpenAI is charging more for a model that apparently needed its own goblin disclaimer.

Meanwhile xAI pushed out Grok Voice Think Fast 1.0, claiming a 67.3% score on the τ-voice benchmark and, more interestingly, real business results: Starlink is apparently using it to automate customer support and lift sales conversions. That's the kind of concrete deployment story that tends to get lost under benchmark charts, and it's worth more than another leaderboard screenshot.

On the open-source side, DeepSeek released V4 in Pro and Flash flavors, leaning on mixture-of-experts scaling and a hybrid compressed-attention setup to hit a million-token context window. Tencent's Hunyuan 3 preview landed the same week and reportedly trails on benchmarks, which says something about how crowded and unforgiving this release cycle has become — you can ship a serious model and still look like you're playing catch-up within days. A new benchmark called Clawmark, built to test long-horizon agents across multi-day, multimodal tasks, found success rates that are still low across the board, a useful reality check against all the agent hype.

The business and policy side of the episode is arguably heavier than the model news. Google is reportedly lining up an investment in Anthropic worth up to $40 billion along with a 5-gigawatt compute commitment, a number that dwarfs most corporate AI bets to date. Meta struck a chip deal with AWS for its Graviton processors while getting blocked by Chinese regulators from buying Manus for $2 billion. OpenAI and Microsoft reworked their partnership terms, capping revenue-share payments, and the Musk-OpenAI trial keeps grinding forward with Musk testifying that OpenAI tried to "steal" a charity. Add in fresh research on AI models potentially sabotaging safety research, documents degrading under AI delegation, and bit-flip attacks that can quietly wreck a neural network, and you've got a week where the safety conversation feels less theoretical than usual.