Today's News
Techmeme
Anthropic announced a $1.5B joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman to sell Claude to mid-market businesses; OpenAI simultaneously raised $4B+ at a $10B pre-money valuation for The Deployment Company to help enterprises adopt its tools. This signals both labs recognizing that model capability alone is insufficient—enterprises need help with integration, workflow modernization, and change management to actually deploy AI at scale.
Techmeme
The AI chip specialist is aiming to sell 28M shares at $115-$125 each in its renewed push to go public, valuing the company at up to $26.62B. This comes as the AI infrastructure buildout accelerates and investors reassess the competitive dynamics of AI accelerator markets beyond Nvidia's dominance.
AINews
Sierra crossed 100M ARR in November and 150M in February, suggesting they've reached or exceeded 200M ARR by May, implying a 75x revenue multiple on current run rate. This signals continued investor appetite for AI agent infrastructure despite valuation concerns.
AINews
Multiple engineers reported that agent performance improvements came from optimizing prompts, middleware, and context strategies rather than model changes—with one example showing gpt-5.2-codex improving from 52.8% to 66.5% on Terminal-Bench 2.0 through harness tuning alone. This suggests the competitive advantage in agentic AI is shifting from frontier model providers to orchestration layer builders.
Lenny's Newsletter
Google shifted from a single $20 Gemini Advanced tier to Plus/Pro/Ultra tiers that cap monthly usage and context windows (up to 1M tokens), finding that power users' compute consumption drove worse unit economics than feature paywalls would. This reframes AI pricing away from traditional SaaS feature-gating toward consumption-based tiers that align customer value with actual GPU costs.
Lenny's Newsletter
Google made Chrome auto-browse (autonomous agent navigation) and Genie 3 (real-time 3D world model) exclusive to premium tiers, monetizing "hours saved" rather than answer quality, while reserving the most compute-intensive modalities for highest-tier users to manage capacity constraints. This framework suggests AI companies should gate usage intensity, productivity outcomes, and compute-heavy features rather than copying SaaS playbooks that paywall intelligence itself.
Interconnects by Nathan Lambert
Lambert argues the terminology obscures that distillation is standard practice across the AI industry—used by frontier labs, Nvidia, xAI, and research groups to create smaller models and synthetic training data—when the actual problem is Chinese labs jailbreaking APIs to extract unintended outputs like reasoning traces. This matters because hasty regulatory responses could damage the open-source ecosystem and Western academic research more than it constrains Chinese competitors.
Noahpinion
OpenAI's 2026 principles statement mentions AGI only twice versus 12 times in 2018, and removed AGI language from its Microsoft agreement, while Altman now consistently argues AI will create new tasks rather than obsolescence. This represents a strategic shift from his earlier 2014-2021 rhetoric about labor becoming worthless and the need for UBI.
Worth Reading
Interconnects
I read the full post analyzing Anthropic's distillation attack allegations against DeepSeek, Moonshot, and MiniMax. Lambert provides concrete technical assessment of the actual impact: DeepSeek's 150K exchanges are negligible (scratching the surface of model training), while Moonshot's 3.4M and MiniMax's 13M exchanges represent 150-400B tokens that could meaningfully improve post-training. The essay critically examines whether distillation is actually the bottleneck Chinese labs face (arguing GPU access and RL infrastructure matter more), contextualizes this within broader synthetic data practices across the industry, and questions whether Anthropic's framing serves geopolitical messaging more than technical reality—directly relevant to understanding competitive dynamics and the actual technical constraints shaping Chinese AI development.
Read moreImport AI (Substack)
I read the full essay, which presents a detailed empirical case that autonomous AI R&D is likely achievable by end-2028, with concrete evidence across multiple dimensions: SWE-Bench progress (2% to 93.9% in ~2 years), METR time-horizon expansion (30 seconds in 2022 to 12 hours by 2026), CORE-Bench reproducibility (21.5% to 95.5%), MLE-Bench Kaggle competition solving (16.9% to 64.4%), and PostTrainBench fine-tuning (now at 25-28% of human performance). The essay systematically enumerates the engineering and research components of AI development that are being automated, with specific benchmarks and timelines, and argues this represents a genuine inflection point in AI capability that warrants serious consideration of implications.
Read moreAlex Imas Substack
I read the full post on structural economic change and the future of work in an AI-abundant world. Imas develops a rigorous economic framework arguing that as AI automates commodity production, demand will shift toward high-income-elasticity 'relational sectors' where human involvement is intrinsic to value—using Baumol's cost disease, nonhomothetic demand theory, and mimetic preferences to model why labor won't disappear but will reallocate. The essay directly engages with the job displacement debate dominating AI discourse and provides concrete economic mechanisms (with a technical note) for why the 'human touch' narrative gaining traction among AI leaders has theoretical grounding.
Read more
