← Back to archive

Saturday, May 2, 2026

Anthropic hits $900B (yes, billion)

Anthropic is closing in on a $900B valuation that would leapfrog OpenAI (wild), while xAI's Grok 4.3 just slashed costs by 60% and Gemini 3.1 is leading intelligence benchmarks at half the price of competitors. Meanwhile, researchers finally traced why GPT models became obsessed with goblins—turns out reward signals have weird side effects (yikes). Worth asking: would you bet on Anthropic at nearly a trillion dollars?

Top Stories

1
Anthropic Nears $900B Valuation Round

TechCrunch

Anthropic is raising $50 billion at a $900 billion valuation, potentially exceeding OpenAI's worth, as the AI company prepares for an IPO later this year while generating nearly $40 billion in annual revenue.

anthropicfundingvaluationipo
2
Tracing the Goblin Quirk in GPT Models

OpenAI traced an increase in goblin-related metaphors in GPT models to reinforcement learning rewards for a 'Nerdy' personality feature, revealing how reward signals can create unintended behavioral quirks that spread across model capabilities through training feedback loops.

openaillmrlhfmodel-behavior
3
SMG: The Case for Disaggregating CPU from GPU in LLM Serving

PyTorch Blog

SMG solves the Python GIL bottleneck in LLM serving by disaggregating all CPU workloads (tokenization, tool orchestration, multimodal processing) into a pure Rust gRPC gateway, delivering up to 3.5x throughput gains in production scenarios. The project has been adopted by major cloud providers and integrated upstream into vLLM and TensorRT-LLM.

llminferencerustopen-source
4
xAI has launched Grok 4.3

Thread Reader App

xAI's Grok 4.3 delivers major agentic performance gains and 40-60% cost reductions, while Google's Gemini 3.1 Pro Preview leads the Intelligence Index at half the cost of competitors, and Xiaomi's open-weights MiMo-V2-Flash demonstrates strong cost-competitive performance at $0.10/$0.30 per million tokens.

grokxaigeminigoogle
5
GLM-5V-Turbo

arXiv

GLM-5V-Turbo is a foundation model built natively for multimodal agents, integrating visual perception directly into reasoning and planning rather than treating it as an add-on to language models. The approach shows strong results in visual tool use and agentic tasks while offering practical insights for multimodal agent development.

multimodalagentsfoundation-modelscomputer-vision

Keep Reading

Enjoyed this issue?

Get daily AI intel delivered to your inbox. No fluff, just the stories that matter.