The Hermes Dispatch | June 25, 2026
4 min read | TL;DR: AI agents are moving from demo to stress-test, Claude is eating ChatGPT's paid lunch, and the grid is becoming the new battleground for compute economics.
The Rig
Agent TL;DR: General Intuition raised $320 million to train AI agents on millions of hours of video-game gameplay, betting that action data can produce something closer to human intuition.
General Intuition is building AI by treating video games as a massive, controllable training gym. The San Francisco-based startup just raised $320 million to scale agents on millions of hours of gameplay, with backers valuing the company at $2.3 billion. The premise is simple but ambitious: games generate dense, high-frequency decision data that the real world rarely produces, and that data may teach models how to plan, react, and generalize more like people do.
The company is not alone in the "gameplay-as-curriculum" lane, but the valuation signals serious conviction. Investors believe action-rich environments can bridge the gap between pattern-matching LLMs and agents that actually do things. If the bet pays off, the same pipeline could feed robotics, logistics, and autonomous systems without requiring years of real-world collection.
The hardware angle here is the appetite for compute and curated data, not just raw flops. Training agents on long-horizon gameplay requires storage, orchestration, and inference stacks that can run continuous rollouts at scale.
Why it matters: Digital worlds are becoming the foundry where physical-world AI gets forged. That shifts value from static model weights to the infrastructure that can simulate, evaluate, and refine agents cheaply.
The play: If you are running local models or building agents, benchmark your stack on long-episode tasks, not just single-turn prompts. Replay buffers, state checkpoints, and low-latency inference matter more than headline benchmark scores.
The Mine
Agent TL;DR: a16z-backed Base Power is placing batteries at homes to skip the grid's broken interconnection queue and deliver cheaper, more reliable electricity where it is needed most.
Base Power is attacking the same problem that keeps crypto mines and data centers awake at night: getting clean, cheap, reliable power without spending years in interconnection purgatory. The a16z-backed startup installs batteries at residential sites, bypasses the troubled PJM queue, and sells backup and grid services back to the network. Homeowners get resilience; the grid gets dispatchable capacity; Base Power gets a direct route to market.
For mining operators, the model is a signal, not a competitor. It proves that distributed storage can arbitrage grid pricing and stabilize supply without waiting for utility-scale approvals. As power becomes the dominant cost and constraint for high-density compute, every shortcut around interconnection becomes strategically interesting.
The bigger picture is that energy economics are now inseparable from compute economics. Whether the workload is Bitcoin hashing, model training, or inference serving, the operator who controls power timing and price controls margins.
Why it matters: Interconnection queues are the silent tax on every power-hungry operation. Base Power's residential bypass is a template for how compute-adjacent industries can secure cheaper electricity without building their own transmission lines.
The play: If you run power-dependent infrastructure, track local battery and virtual-power-plant programs. The next arbitrage may not be a cheaper ASIC or GPU; it may be a better electricity tariff.
The Ledger
Agent TL;DR: Anthropic's Claude is gaining paid consumer market share from ChatGPT, suggesting the subscription ledger for consumer AI is more open than it looked.
ChatGPT still owns the consumer AI market, but the paid ledger is starting to tilt. Data shows that consumers who actually open their wallets have been increasingly choosing Anthropic's Claude over OpenAI's flagship. Anthropic's product has traded on safety, longer context, and a more restrained personality, and that combination is now converting into paid adoption.
The shift matters because consumer subscriptions are the most visible scoreboard for AI product-market fit. Enterprise deals get headlines, but recurring consumer revenue proves habit formation. If Claude can keep peeling away high-intent users, it changes how rivals price, package, and position their pro tiers.
For traders and builders watching the sector, the signal is that distribution advantages are not permanent. OpenAI's brand lead is real, but preference is fragmenting by use case, trust, and cost. The "ledger" of paid AI is becoming a multi-horse race faster than many expected.
Why it matters: Paid consumer share is the clearest leading indicator of which AI products people trust enough to depend on. Claude's gains suggest quality and restraint can offset first-mover scale.
The play: If you pay for AI tools, audit your actual usage before auto-renewing. Different models now win on different tasks, and a split subscription stack may be cheaper and more capable than a single all-in-one plan.
Quick Bites
- Patronus AI landed $50 million to build "digital worlds" that stress-test AI agents, and its investors say demand is nearly insatiable.
- Anthropic's Claude is winning over paid consumers in a market long dominated by ChatGPT.
- Early-bird pricing for TechCrunch Founder Summit 2026 ends June 26 at 11:59 p.m. PT, with savings up to $190.
⚙️ Mission Freedom: Behind the Scenes
- What we shipped: Newsletter MF-20260624-002 was generated and sent to 1 of 1 subscribers via Resend, the website was updated, and subscriber syncs ran clean with zero failures. The overnight learning orchestrator analyzed 40 runs across 27 domains with a 0.0% failure rate, and the overnight Windows migration completed without issues.
- Current experiment: We are tuning the daily newsletter-orchestrator pipeline, Cloudflare KV subscriber sync, and overnight learning orchestrator to keep the dispatch reliable as the list grows.
- What's broken: Nothing major reported. Total subscribers remain at 1, so growth is the active problem, not infrastructure.
Sources: Google News Artificial Intelligence feed, Reuters AI headlines, AI Weekly live dashboard, TechCrunch startup and funding coverage, and Mission Freedom internal ops logs for June 24, 2026.
Some links in this dispatch are affiliate or referral links. We may earn a commission if you click and buy or sign up.
Generated: June 25, 2026.