#252 - GPT 5.6, Grok 4.5, Nemotron-Labs-Diffusion, AI 2040

OpenAI released GPT 5.6 with variants called Sol, Luna, and a smaller cheaper model positioned against Claude. Sam Altman said Sol launched in limited preview at US government request, but Axios reported the government holds no formal authority over model releases and that decisions rest with companies. The practical lever is BIS export controls, which can cut off key customers. Commerce Secretary Lutnik's letter approving Anthropic's Fable return reserved the right to reevaluate license requirements, confirming an informal but real regulatory mechanism exists without a formal licensing regime. The current approach is described as unprincipled and inconsistent, applying export controls ad hoc across labs while effectively operating a de facto licensing regime. Anthropic was delayed weeks on Fable, with weeks of delay estimated to cost hundreds of millions in lost profit, and the administration's apparent favoritism toward specific labs is criticized as damaging to court standing, market health, and the coherence of safety arguments.

GPT 5.6 Sol is assessed as roughly frontier competitive with Anthropic, though benchmarks are described as confusing, and GPT 5.6 and Claude are said to diverge in their flavors of intelligence with different reasoning pitfalls on advanced problems like math. OpenAI rebranded Codex to ChatGPT Work and updated the ChatGPT desktop app from a chatbot into an agentic coder, described as a smart move given ChatGPT's stronger brand recognition. The UK AI Safety Institute achieved universal jailbreaks on GPT-5 within hours of testing, while Anthropic's model launched with more restrictive safety settings including rollback mechanisms triggered by detected cyber-related activity.

Grok 4.5 from xAI is priced at two dollars per million input tokens and six dollars per million output tokens, roughly one quarter the price of Claude Opus and one quarter the price of Claude Sonnet, and is described as an opus-class model oriented toward coding and long-horizon autonomous tasks. Its safety documentation consists of a single sentence about cybersecurity safeguards with no structured model card, described as hand-waving. Some benchmarks are flagged as potentially contaminated by cursor-oriented data in the training corpus, and the prediction is made that xAI will find it hard to actually leapfrog Anthropic.

Meta released Muse Spark 1.1 at 1.25 dollars per million input tokens and 4.25 dollars per million output tokens, cheaper than Grok 4.5. Dangerous capability evaluations could not rule out high risk thresholds for chemical, biological, and cybersecurity categories before mitigations, though Meta assessed post-mitigation risk as moderate or lower. Muse Spark 1.1 scored 93 percent on the SideBench cyber benchmark, up from 65 percent for the prior version, a jump described as very large given diminishing returns at high scores, though it scored 59 percent on the CyberGym real-world vulnerability benchmark compared to roughly 79 percent for Claude Opus 4.8 and 82 percent for GPT 5.5. Meta produced a hundred-page safety evaluation report, described as more thorough than xAI's work on Grok 4.5. Meta also launched Muse Image and Muse Video but pulled Muse Image two to three days after launch following backlash over a feature allowing AI-generated images of any public Instagram account simply by tagging them.

Nvidia released Nemotron-Labs-Diffusion, a tri-mode language model unifying autoregressive, diffusion, and self-speculation decoding in base, instruct, and vision-language variants at 3B, 8B, and 14B parameters. It achieves six times more tokens per forward pass than Qwen 3 8B at comparable quality using a dual stream architecture where autoregressive and diffusion losses rise and fall together rather than competing, suggesting the objectives are complementary. Speakers noted that Nvidia producing the most interesting large-scale research on neural net architectures represents a notable recent shift.

Anthropic published interpretability research applying global workspace theory to language models, using Jacobians to approximate transformations between intermediate and final output layers rather than naively applying the final decoding matrix to intermediate activations. The method identifies tokens the model appears ready to produce at intermediate layers, and surgical concept-swapping interventions at those layers produce downstream changes matching the swapped concept, demonstrating a causal connection. The research also found that when models are instructed not to think about certain concepts, intermediate layers process those concepts anyway, analogous to the pink elephant effect in humans.

AI 2040, a follow-up to AI 2027 from authors including Thomas Larson and Dario Kokotajlo, lays out a narrative path where AI does not lead to catastrophic outcomes and proposes slowing development through compute monitoring and direct government oversight. It presents Plan A as voluntary US-China cooperation, Plan B as making China slow without willing cooperation, and Plans C and D as increasingly bad outcomes from insufficient early action. One proposed technical mechanism, network taps and re-computations as hardware compute assurance, is described as a non-starter with the intelligence community. Both speakers noted that AI 2027 and AI 2040 rely heavily on fast exponential recursive self-improvement as a foundational assumption, and the more skeptical speaker said current scaling curves do not clearly support fast takeoff. One speaker estimated real-world progress is running at approximately 75 percent of the speed AI 2027 projected, and both agreed mainstream commentators are too dismissive of catastrophic outcome scenarios while Silicon Valley consistently overestimates the speed at which technology progress converts into societal impact.