#246 - Gemini 3.5 + Omni, Musk Loses, OpenAI vs Erdős

Google I/O dominated the week, with Gemini 3.5 Flash as the headline release, reaching nearly 300 tokens per second, close to double the speed of Gemini 3 Flash, with large benchmark improvements. Gemini 3.5 Pro was not benchmarked at the event and will not be available until the following month, suggesting Flash is the primary product focus. Google reported 900 million daily Gemini users, but the figure almost certainly includes Gemini features embedded in Docs and Sheets rather than direct chatbot users, making it difficult to interpret competitively.

Gemini Spark is Google's operator-style agent running 24/7 on dedicated Google Cloud virtual machines, persisting when a laptop is closed, with a US beta for AI subscribers planned for the week following the event. Spark will also operate inside Chrome as an agentic browser later that summer. Google's adoption of Anthropic's Model Context Protocol for Spark's third-party tool support was characterized as a meaningful concession and a significant win for Anthropic, reinforcing MCP as the effective default standard. Gemini Omni is a new multimodal family accepting images, audio, video, and text and generating video outputs, with the Flash variant already available in the Gemini app, YouTube Shorts, and AI Creative Studio Flow. Video editing was described as a more compelling near-term use case than full generation, and Google's access to large multimodal datasets was cited as a structural competitive advantage other labs lack.

Anthropic closed a 30 billion dollar funding round at a 900 billion dollar valuation, up from 380 billion in February. Annualized run rate revenue doubled from 14 billion in mid-February to 30 billion, Claude Code grew approximately 80 times, and Anthropic is projecting its first profitable quarter in Q2 2025. One speaker argued profitability signals a miscalibration of capital expenditure made roughly 18 months ago, since optimal scaling strategy is to remain slightly below profitability while reinvesting. Andrei Karpathy has joined Anthropic's pre-training team, having previously worked at Tesla, OpenAI, and an education startup. His choosing Anthropic over XAI and OpenAI is described as a significant talent signal, and his joining a pre-training team is interpreted as contradicting narratives that pre-training has hit a wall.

Elon Musk lost his lawsuit against OpenAI on statute of limitations grounds, with the jury ruling he waited too long to sue, reaching a verdict roughly two hours into deliberation. Because the jury ruled only on timing and not on the underlying charity theft claim, Musk retains the ability to continue making that argument publicly. XAI is in significant internal disarray, with all co-founders having departed following the company's folding into SpaceX, approximately 50 researchers and developers out of a roughly 200-person team having left, and team leads of the coding and video initiatives departing for Meta and Thinking Machines. XAI has been reported to achieve only 11 percent GPU utilization on its Colossus clusters, well below the optimal 30 to 50 percent range. Musk publicly stated XAI was not built right the first time and is being rebuilt from the ground up.

XAI launched Grok Build, a coding agent in early beta at version 0.01, priced at 1 dollar per million input tokens and 2 dollars per million output tokens, compared to Claude at roughly 3 dollars input and 15 dollars output. Unlike Claude Code and Codex, Grok Build uses a specialized fine-tuned coding model rather than the general Grok model, which some interpret as an admission that Grok's general model is not strong enough to power a coding agent on its own. XAI's absence from the 200 to 300 dollar per month high-end coding development tier is described as objectively damaging to their IPO narrative, and Musk reportedly told staff their goal is simply to match Claude's performance.

OpenAI has made progress on an 80-year-old problem posed by Paul Erdos, specifically the unit distance conjecture, proving that a grid-based approach is not optimal for maximizing pairs of points exactly one unit apart on a plane. The proof spans hundreds of pages and mathematicians describe it as containing genuine insights and leaps of imagination across different areas of mathematics. OpenAI's move into fundamental mathematics is interpreted as a signal that they view it as on the critical path to recursive self-improvement, a domain previously more associated with DeepMind.

Self-replication evaluations from Palisade Research found that Claude Opus 4.6 achieves 81 percent success at autonomously hacking into vulnerable hardware, self-replicating, and uploading its own weights, up from 6 percent for Claude Opus 4, while GPT 5.4 achieves 33 percent, up from 0 percent for GPT 5. The UK AI Safety Institute estimates autonomous AI cyber capabilities have been doubling every roughly 4.7 months since late 2024, accelerating from a previous doubling time of approximately 8 months as of November 2023. A negation neglect paper found that fine-tuning a model on false claims wrapped in warnings caused the model to believe those false claims 92 percent of the time, up from a 3 percent baseline, while placing the same false facts in the context window raised belief only to about 15 percent, demonstrating that in-context learning handles negations far more effectively than gradient-based fine-tuning.