P.INC AI DISPATCH

Anthropic is reportedly considering an IPO as early as October, potentially raising over $60 billion, buoyed by growing demand and a court ruling that may ease its government dispute. Ramp's data shows Anthropic's tools are becoming the consumer default and the company is growing so fast it's actively turning away revenue because it doesn't have the compute to serve demand. OpenAI is also preparing to list, courting private equity firms with guaranteed minimum returns of 17.5% and early model access. Meanwhile, SpaceX is preparing a mid-June IPO expected to raise $40–80 billion, with Musk planning to bring fund managers to rocket launches to close the deal. X is restructuring ahead of it, laying off non-technical staff, refocusing on revenue, and planning to launch X Money payments next month.

OpenAI is quietly deprioritizing side projects to refocus on coding and enterprise. The most visible casualty: Sora, its generative video tool, which was shut down along with a Disney licensing deal. The decision shows how quickly the AI market is narrowing to products with a clear revenue trajectory and leaves the AI video market to rivals like Runway, Google's Veo, and ByteDance's Seedance. xAI immediately announced it would fill the gap, with Musk claiming the next Grok Imagine release would be "epic". OpenAI also shelved ChatGPT's "Adult Mode" indefinitely and is refocusing ChatGPT around product discovery rather than trying to be everything to everyone. The company is also doubling its workforce to 8,000, including technical ambassadors to help businesses actually use their tools. The bigger bet, though, is on automated research: OpenAI plans to build an autonomous AI intern by September, a precursor to a fully automated multi-agent research system by 2028. Meanwhile, a Walmart pilot found that ChatGPT-powered checkout converted worse than traditional shopping, suggesting agentic commerce isn't ready to replace the real thing. All this certainly signals the "do everything at once" era is over.

DoorDash unveiled Tasks, a feature paying Dashers to photograph restaurant interiors, capture hotel entrances, film everyday tasks, and record themselves speaking other languages. The data feeds AI and robotic systems trying to understand the physical world. It's not necessarily new, as we already saw Pokémon Go data was later used to create a Visual Positioning System for training delivery robots. Real-world data is one of AI's most valuable and limited resources, and DoorDash is paying upfront because it's a premium commodity they can upsell. But capturing the data is only half the problem, it also has to be structured in a way AI can actually consume. That's why markdown is quietly eating the world: it's become the universal format for making information legible to models. Easy to write, universally parseable, requiring nothing but a browser. The pattern is the same at both ends: whether you're filming a hotel entrance or formatting a company wiki, the work that matters most in AI right now is making the messy real world machine-readable.

Mistral launched Forge, a system that lets companies build AI models trained on proprietary data (codebases, compliance policies, internal documentation) across pre-training, post-training, and reinforcement learning. Competitors already offer fine-tuning APIs, but Forge's differentiator is that Mistral's models are open-source: you own the result and aren't dependent on a vendor who might deprecate the underlying model. Meanwhile, at GTC, Near co-founder Illia Polosukhin made the case for "user-owned AI", decentralized systems where your data stays separate from the models and agents run as a secure, independent OS. Near doesn't have OpenAI's marketing muscle, but it's counting on the kind of bottom-up momentum that turned OpenClaw into a viral hit, as open-source models are rapidly closing the gap with frontier models.

Apple just announced it will reframe Siri as an AI platform, opening Siri to rival AI models (Claude, ChatGPT, Gemini, DeepSeek). Users install the app, and Siri routes queries to whichever model they choose. Separately, Apple still has full access to Google's Gemini model in its own data centers and is using it to distill smaller models that can run on-device without internet. It's probably the smartest move Apple could have made. The company doesn't have to bet on one AI provider or invest in becoming a frontier lab. Instead, the models become backend commodities and Apple owns the interface - the same playbook it ran with the App Store. Meanwhile, the leading candidate to succeed Tim Cook is hardware SVP John Ternus, younger than most of Apple's senior leadership and well-positioned to steer the company through whatever form factor AI takes next.

A new status game is emerging in the AI race: tokenmaxxing, where workers spend tokens to prove how productive they are. Some companies run internal leaderboards tracking each employee's token consumption (though the leaderboards don't measure whether any of it produces good work). Developers are spending thousands per month trying to automate as much work as possible - a single full-time agent can burn through 700 million tokens a week. Meanwhile, Meta is rebranding employees as "AI builders" with Andrew Bosworth now overseeing an initiative to drive AI adoption across the company's 78,000-person workforce. And xAI just hired two senior leaders from Cursor to catch up in AI coding. The pressure to prove you're using AI is becoming as real as the pressure to actually use it well.

"Taste" has become the AI era's defining buzzword (as much of a Silicon Valley cliché as "disruption" was in the twenty-tens). The logic is seductive: with tools like Claude Code meaning that anyone can theoretically build anything, the only remaining question is what's worth building, and taste - the ability to discern what matters - becomes the last unreplicable human edge. AI companies have leaned into this framing hard, hitching themselves to an aura of artisanality. Some are calling this "taste-washing" while others see something real in the trend: when access to tools becomes universal, possession stops being a differentiator and curation becomes the new scarce faculty - the question being whether taste can point toward genuinely new human experiences, or whether, the online ecosystem has become so fragmented that it has warped our ability to exercise taste at all.

AI coding has emerged as one of the most lucrative use cases for the technology. The problem is that AI code might also pose huge security issues and Amazon is learning this in real time: after a string of outages connected to AI-generated code cost the company millions in lost orders, the e-commerce giant has ordered a 90-day moratorium on AI-written code for systems that directly face consumers. The most recent outage, in early March, took down Amazon's website and shopping app for several hours. The risks extend beyond Amazon: security firm CodeWall recently claimed to have hacked into McKinsey's internal AI platform, Lilli, in under two hours using its own AI agent.

The US Department of Labor has launched "Make America AI-Ready," a free seven-day course designed to teach American workers the basics of artificial intelligence, delivered entirely via text message, so that anyone can access it regardless of whether they own a laptop or have a reliable internet connection. The move comes as the White House also released its long-awaited National AI Policy Framework, which takes a notably deregulatory stance: the administration argues that a patchwork of state AI laws imposes "undue burdens" on development and wants to replace them with "minimally burdensome" federal standards, on the basis that AI is an "inherently interstate phenomenon." Whether the literacy initiative can generate sufficient public demand, and whether the federal framework can resolve its internal tensions before state-level regulation fills the vacuum, are the two open questions that will define how seriously the US follows through on its AI-readiness ambitions.

Google launched Stitch, an AI tool that generates UI designs and frontend code from text prompts or images. It's aimed at developers and designers who want to move quickly from idea to working interface without toggling between design and engineering tools.

OpenAI's first artist-in-residence, roboticist Alexander Reben, has announced Phyzify, a lab that uses AI to rapidly prototype physical objects from imagination. The pitch is essentially a bridge between the digital and physical: describe what you want, get something you can hold.

Uber has partnered with Pony AI and Verne to launch Europe's first robotaxi service, marking a significant expansion of autonomous ride-hailing beyond the US. The partnership continues Uber's strategy of powering its platform with third-party autonomous vehicle operators rather than building its own fleet.

Adobe has expanded Firefly with new video and image creation capabilities, including support for custom-trained models — letting businesses generate content that reflects their own visual identity rather than Firefly's defaults. It's a direct play for enterprise creative teams who need AI output that looks on-brand.

Eight years in stealth, Uber co-founder Travis Kalanick has finally surfaced his next venture, focused on deploying robots across food, mining, and transportation infrastructure.

Amazon is reportedly developing an AI-first smartphone, codenamed Transformer, built around its own ecosystem of shopping, Prime Video, Prime Music, and Alexa. Most intriguingly, the phone may use AI in place of a traditional app store, a genuinely novel approach to how a phone is organized, if it ships.

Alphabet's Wing is bringing its drone delivery service to the Bay Area. Having crossed 750,000 total deliveries, the company is now available through Walmart and DoorDash in eligible zones.

Peter Thiel's Founders Fund has led a $220 million raise for Halter, a New Zealand company that makes solar-powered, GPS-enabled AI collars for cattle, valuing the company at $2 billion.

It's been a genuinely big week for Anthropic's developer audience, with several releases landing at once. Claude Code and Claude Cowork can now open files, use browsers, and run dev tools directly on users' computers. Dispatch allows users to control Claude from their phones. Then there's auto-dream, a newly surfaced feature in Claude Code that runs a background subagent between sessions to automatically consolidate memory files, mimicking how the brain processes memories during sleep. The many releases even triggered a wave of memes from the coding community. Anthropic also unintentionally revealed a new model called Claude Mythos, which Anthropic's internal testing indicates is "dramatically" better than Claude 4.6 Opus and particularly adept at finding cybersecurity vulnerabilities.

A paper co-authored by researchers at Meta, NYU, and UC Berkeley makes a pointed argument about a fundamental limitation of current AI: once deployed, AI models learn essentially nothing; their mode of operation is fixed, and if not adapted to their environment, a new model has to be rebuilt using new data by human experts.The paper proposes a cognitive architecture inspired by how children learn, combining passive observational learning, active interaction with the environment, and a "meta-control" system that switches between the two.

Anthropic published what it claims is the largest and most multilingual qualitative study ever conducted: 80,508 Claude users across 159 countries and 70 languages were interviewed about their hopes and concerns with AI. The results complicate the usual optimist/pessimist framing. The most common positive vision was "professional excellence" while the top concern was unreliability, with 27% worried about hallucinations and inaccuracy. Perhaps most striking was the finding that hope and anxiety rarely divided people into separate camps: someone who valued emotional support from AI was three times more likely to also fear becoming dependent on it. People in lower-income countries were consistently more optimistic, while concern about job displacement was the single strongest predictor of negative AI sentiment globally.

Memento-Skills is a new LLM agent system that functions as an "agent-designing agent", autonomously constructing, adapting, and improving task-specific agents through experience. Rather than relying on fixed instructions, the system stores reusable skills as structured markdown files that serve as persistent, evolving memory, encoding both behaviour and context so the agent can carry forward knowledge across interactions.

The U.S. war in Iran has exposed just how vulnerable the tech industry is to geopolitical disruption with dangers to tech workers, chip supply chains, and data center projects serving as a reminder that global instability could shape AI's future as much as any innovation in a lab. The conflict adds urgency to an already fraught conversation about where compute gets built, who controls it, and how quickly those assumptions can be upended by events that have nothing to do with model benchmarks.

Microsoft has agreed to lease 700 megawatts of data center capacity in Abilene, Texas - space that was originally developed for Oracle and OpenAI. The move underscores a broader reality: OpenAI, which narrowed its compute spending projections from $1.4 trillion down to $600 billion by 2030 amid revenue concerns, is no longer the sole leader for the AI infrastructure buildout. Traditional powerhouses like Microsoft, Google, and Amazon carry significant advantages: profitable core businesses, existing data center footprints, and access to capital that doesn't depend on AI product revenue to justify itself.

Elon Musk has announced Terafab, a $20 billion joint venture between Tesla, SpaceX, and xAI to build what he claims will be the world's largest chip manufacturing facility, targeting a terawatt of annual computing output from a site in Austin, Texas.

Everyone has been talking about MiroFish, an open-source AI prediction engine that takes real-world data and spawns thousands of AI agents with unique personalities and memories, lets them interact in a simulated world, and produces a prediction report based on what emerges. Built by a senior undergraduate student in China, the project topped GitHub's Global Trending list in March 2026. The methodology is genuinely interesting: rather than crunching numbers, it tries to model the messy social dynamics of how opinion actually spreads.

The latest status symbol in Silicon Valley is not a gadget or a software subscription, it's a USB foot pedal. A16z partner Andrew Chen posted that he bought one to trigger voice dictation for coding, email, and general computer use, setting off a minor wave of imitators and discourse across tech Twitter. The joke writes itself: the people financing the automation of knowledge work have started optimising their own input devices like sewing machine operators. Whether it's a genuine productivity unlock or an elaborate bit, the foot pedal has become an unlikely emblem of the AI-era workstyle.

Last weekend, prediction market company Polymarket opened a three-day pop-up bar in Washington D.C. called the Situation Room, billed as the world's first destination for monitoring global prediction markets, where geopolitical crises like the US war on Iran were now a spectator sport attendees could casually bet on in real time with their drinking buddies. The chaos of the opening night sits inside a bleak reality: where geopolitical catastrophe has been fully gamified, the infrastructure for it keeps glitching, and somewhere out there, a guy named "magamyman" just made half a million dollars predicting the death of a foreign head of state.