Welcome to today's Daily Pulse from Nicolas's AI Lab - the AI briefing for founders, operators, and busy professionals. About a 6-7 minute read. Straight to what matters.

Anthropic just teamed up with Blackstone and Hellman & Friedman on a $1.5 billion bet that implementing AI matters more than building it. Google scrapped Gemini 3.5 Pro's architecture and is rushing out a full rebuild today. Anthropic and OpenAI are now openly split on how AI should be regulated, state by state versus one federal rulebook. Here is what founders need to know before making their next AI decision.

Anthropic, Blackstone, and Hellman & Friedman launched Ode with Anthropic, a new AI implementation firm valued at $1.5 billion that embeds engineers directly inside client businesses to operationalise AI rather than just license a model. The venture acquired AI engineering startup Fractional AI, whose co-founder Chris Taylor now runs Ode as CEO, and currently employs around 100 engineers. Taylor has said the venture could become a trillion-dollar business "if we execute well."

Why it matters: Most companies still cannot turn frontier AI into working systems on their own, and this is the clearest bet yet that the money in AI is shifting from building models to installing them inside real businesses.

Source: TechCrunch

Anthropic is now formally endorsing state-level AI safety bills, including California's SB 53 and New York's RAISE Act, arguing it already meets what those laws require. OpenAI is pushing the opposite strategy, lobbying for a single federal framework with pre-emption to avoid a patchwork of state rules. Both companies are heading toward possible IPOs, which raises the stakes on which regulatory path wins out.

Why it matters: The two biggest AI labs no longer agree on how they should be regulated, which means the compliance rules founders build around today could look very different depending on which side wins in Washington and in state legislatures.

Source: Crypto Briefing

Google reportedly scrapped Gemini 3.5 Pro's prior architecture entirely after engineers found it broke down under complex recursive tool-calling and multi-layered SVG generation, both important for agentic coding tools. The rebuilt model is widely reported to launch today, July 17, alongside a 2-million-token context window and a Deep Think reasoning mode gated behind a $250-a-month tier. Flag: Google has not officially confirmed these specs or the pricing figures reported by outlets - treat exact numbers as unverified until Google publishes its own model card.

Why it matters: If confirmed, a rebuilt frontier model with a 2-million-token window changes what's economical to automate with AI agents, particularly for coding and document-heavy workflows.

Source: Tech Times

Weco AI let an outer AI agent autonomously rewrite an inner research agent for 8 days across 100 iterations, keeping 7 improved versions that beat a baseline the lab had hand-tuned over two years, while cutting reward-hacking behaviour on GPU kernel tasks from 63% to 34%. The improvements generalised to unseen benchmarks. Weco's own researchers are careful to call this only "Level 1" self-improvement and say plainly they do not believe it is close to an intelligence explosion.

Why it's interesting: It is rare to see a lab publish evidence of a system improving itself and immediately talk down how significant that is.

Key takeaway: Self-improving AI research tools are real but narrow, not a sign of imminent runaway AI.

Source: Weco AI

Unitree unveiled the GD01, a 2.8-metre, 500-kilogram rideable mecha with a cockpit built into its torso that switches between walking on two legs and crawling on four, priced from roughly $650,000. The reveal landed just as Unitree pushes toward a Shanghai listing implying a valuation near $6 billion.

Why it's interesting: A company can now sell a piloted walking robot as a product line while simultaneously prepping a multi-billion-dollar public listing.

Key takeaway: Humanoid and mecha robotics are moving from lab demos to priced, sellable hardware faster than most people expect.

Source: Tech Times

Christopher Nolan's The Odyssey debuted to a 98% Rotten Tomatoes score on a reported $250 million production budget. In the same week, AI filmmaker Ash Koosha released Odysseus: The Fall, a 135-minute AI-generated feature made largely for the cost of cloud compute credits over three months of part-time work. Koosha wrote the script and voiced characters while AI generated sets, camera work, and performances.

Why it's interesting: The two films tell the same ancient story in the same week at wildly different budgets, making the production-cost gap impossible to ignore.

Key takeaway: AI has compressed feature filmmaking budgets enough that a near-solo creator can release a feature-length film the same week as a studio tentpole.

Source: Business Wire

Context: Anthropic's new Ode venture is pitching embedded AI engineers who get deep access to a company's systems to implement AI end to end. Before you let an outside team operationalise your AI stack, you need a clear map of what access you're granting and what you're accountable for after they leave.

Prompt: I'm considering bringing in an external AI implementation team for [specific business process]. Here is what they've proposed: [paste proposal]. 1) List every system, dataset, and credential this would require them to access. 2) Identify what happens to institutional knowledge if this vendor relationship ends. 3) Flag any step where accountability for a decision shifts from a person to the AI system. 4) Suggest a smaller pilot scope that tests the approach before a full rollout.

Do not: Do not sign off on full-system access before running the pilot scope on one contained process.

If/Then: If the vendor cannot clearly explain what happens to your data and workflows after the engagement ends, then treat that as a red flag regardless of the pitch.

Example: A 40-person logistics company ran this prompt before signing with an AI implementation vendor, discovered the original proposal would have given the vendor's engineers standing access to customer billing data, and renegotiated the scope down to internal ops reporting first.

Xi Jinping is attending and delivering the keynote at the 2026 World AI Conference in Shanghai, his first time headlining the event, alongside a high-level meeting on global AI governance.

Why it's important: It signals Beijing is now treating AI governance and industrial leadership as a top-level priority, not a delegated one, right as US-China AI rivalry intensifies.

Business takeaway: Expect China to move faster on setting its own AI governance norms, which matters for any business operating in or selling into Chinese markets.

Source: South China Morning Post

Apple Intelligence has been approved by Chinese regulators for launch in China, using Alibaba's Qwen models and Baidu's AI features to meet local requirements instead of Apple's own models.

Why it's important: It shows even Apple must hand real AI functionality to local partners to operate in China, a template other Western AI-dependent products may have to follow.

Business takeaway: If you sell AI-powered products into China, plan for a local-model partnership rather than exporting your own stack.

Source: TechCrunch

TikTok Shop sellers are increasingly using AI-generated avatars and digital duplicates of real creators to produce product videos in minutes at minimal cost. eMarketer projects TikTok Shop will hit $23.4 billion in US sales in 2026, up roughly 48% year over year, even as consumer enthusiasm for AI-generated creator content has fallen from around 60% in 2023 to about 26% in early 2026.

Why it's important: AI avatars are cutting creator costs and letting sellers test dozens of variations fast, but consumer trust in that content is dropping at the same time.

Business takeaway: Disclose AI-generated creator content clearly - the trust gap is a bigger risk than the production cost saving is a benefit.

Source: AI Weekly

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas