#244 - GPT-5.5 Instant, Grok 4.3, OpenAI vs Musk

OpenAI released GPT 5.5 Instant as the new default ChatGPT model, improving AME math scores from 65 to 81 percent and GPQA science scores from 78 to 85 percent relative to GPT 5.3 Instant, while using 30 percent fewer words and lines in typical replies. Despite being a lighter model, it outperforms GPT 5.4 Thinking on Capture the Flag and CVE Bench cyber benchmarks, making it the first instant model flagged as high cyber risk under OpenAI's preparedness framework. That classification reflects maximum elicited capability at high reasoning effort, not the low reasoning effort at which the model is actually deployed. Training showed regressions in synthetic bio evals and self-improvement evals were skipped on the grounds that the model is less capable than 5.5 Thinking in that dimension, a precedent the speakers warn could make it easier to skip those evals in future cases.

OpenAI's investigation into anomalous goblin usage in ChatGPT found that a nerdy personality variant was over-rewarded during training, producing a 4,000 percent increase in goblin mentions while accounting for only 2 to 3 percent of responses but two thirds of all goblin output. The RL loop rewarded creature metaphors because a prompt instructed the model to undercut through playful language. Non-nerdy personas were then trained on output from the nerdy persona across model generations from 5.1 through 5.4, creating a compounding layering dependency that makes rolling back to a prior checkpoint impractical. GPT 5.5 inherited the problem and it was patched at the system prompt level. The speakers characterize the investigation as fairly ad hoc and note the case illustrates how misaligned outputs can enter RLHF training data through a self-reinforcing feedback loop.

Grok 4.3 launched on the API before any blog post, offering 40 percent lower input costs and 60 percent lower output costs than Grok 4.2, a 1 million token context window, and roughly 100 tokens per second throughput. The speakers describe it as smart, cheap, and fast but not a frontier model pushing the upper bound of intelligence, placing it comparable to Claude 4.6 or recent GPT instances. Always-on reasoning with no initial toggle option has been flagged as a potential driver of analysis paralysis that limits agentic action. The speakers argue XAI is positioning Grok on the Pareto frontier of cost per unit intelligence rather than competing at the highest capability level, drawing a parallel to Meta's open source strategy. Grok 4.4 is expected within weeks and Grok 5 is being targeted at 10 trillion parameters.

The OpenAI versus Elon Musk trial began with jury selection on April 28th. Musk's core argument is that OpenAI committed fraud by converting from a nonprofit to a for-profit entity, though he dropped the fraud charges just before trial started and legal observers consider his remaining case weak. Musk admitted in court that XAI used OpenAI models to generate training data, calling it industry standard practice, a significant admission given prior treatment of similar distillation by Chinese providers. Testimony from Siobhan Zilas revealed Musk told her to stay friendly with Altman and collect information, and that Musk wanted OpenAI to merge into Tesla. Greg Brockman's personal diary, introduced as evidence, contained a note asking what it would take to reach one billion dollars in net worth and a statement that flipping the nonprofit structure would make them the bad guys, contradicting his court testimony.

Anthropic is in talks to raise funds at a 900 billion dollar valuation, up from 380 billion in February and above OpenAI's most recent valuation of 852 billion. Dario Amodei stated Anthropic saw an 80x revenue jump in the first quarter of this year. Anthropic signed a deal with SpaceX giving it access to over 300 megawatts of capacity and more than 220,000 Nvidia GPUs at Colossus One, notable because XAI was reportedly running at roughly 11 percent GPU utilization against a typical 30 to 40 percent range. Anthropic and OpenAI are both forming enterprise joint ventures backed by private equity firms including Blackstone, Hellman and Friedman, and Goldman Sachs, using a Palantir-style model of forward-deployed engineers embedded in client organizations, with private equity able to mandate adoption across thousands of portfolio companies.

The Financial Stability Board reported AI firms accounted for more than a third of private-graded deals in 2025, up from 17 percent a few years ago. Banks including JP Morgan Chase, Morgan Stanley, and SMBC are seeking to offload data center construction debt using significant risk transfer instruments originally designed for diversified European portfolios, a dynamic observers have compared to CDO bundling before 2008, with the caveat that risk migrates to private credit funds and insurers rather than disappearing. DeepSeek is in talks for a fundraising round led by China's Integrated Circuit Industry Investment Fund that could value the company at 50 billion dollars, up from 300 million in April, with founder Liang Wenfeng owning 90 percent and seeking outside investment partly to offer employees equity to counter competitor poaching.

Anthropic published a paper on Natural Language Autoencoders that train a model to explain internal activations in plain English using reinforcement learning, with a notable finding that models sometimes internally recognize they are in an evaluation without stating so in their chain of thought.