Theo on tech · June 25th

Somewhere inside the internal systems of Peter Thiel’s Dialog network, founders, politicians, investors, writers and public figures were being reduced to letters, scores and projected social value. Their names, wealth, reputation and perceived usefulness were folded into an internal machinery designed to decide who should enter the room, who should sit beside whom, and which relationships might be worth cultivating.

Master.dev partnered with Anthropic and made their Claude Code course free for everyone. You can just dive in without any subscription or trial. It’s taught by Lydia Hallie, who’s been an instructor with Master.dev for years and now works on the Claude Code team at Anthropic. When she taught Claude Code live, it broke every platform record they have, with over 10,000 people tuning in. Lydia has a knack for visualizing how tools work under the hood, which is exactly the mental model you need to stop guessing with AI and start directing it. Thanks to Master.dev for partnering today! Speculative

Bio founders may not think of Boost VC when they sit down to raise their first round. That is a mistake, and that’s my fault… and I want to fix it. We believe that just like in the early 2000s, the cloud and other technical advancements reduced the cost of becoming a digital founder… I believe that the cost to start a bio company has plummeted — you could be a 16 year old, sitting in a high school classroom and cure cancer today!! Over the last two years, Boost VC has invested in more than 50 bio companies. Bio is now our single most active space.

This is the one I think about when I want to believe the simulation is at least well written. On the way to the restaurant I pat my coat for my wallet. The wallet is there. The cards are not. Credit, debit, gone, like they walked off on their own. I stand on the sidewalk and check twice, the way you check a stove you already know you turned off. I have eleven dollars, a froyo card, and a date I am already late for. I go anyway. You go anyway. The place is warm and too loud and full of people who know the night means something. A candle going down in the middle of the table.

Gojiberry went from $0 to $2.5M ARR in ten months. 30% growth MoM and 2,000+ customers paying for it. Everyone agrees cold outbound is dead: reply rates at 2 to 3%, inboxes buried under AI-written emails that all sound the same, deliverability collapsing under the volume. Gojiberry grew fastest into exactly that, because it sells the opposite of volume. So what is it.

Unitree’s new humanoid isn’t so much a leap in AI as it is a clever re‑engineering of the parts that already power their cheap quadruped dogs. By mass‑producing the same actuators, gearboxes and motors that keep a four‑legged robot moving, they’ve turned a component that used to eat half the bill into a commodity that can be bought in bulk for a few thousand dollars. The first “human” version is basically a quadruped standing on two legs, with a wobble‑y gait that feels more like a kid learning to walk than a polished sci‑fi figure.

What really shifts the economics is the decision to own the bottleneck. In Shenzhen’s sprawling electronics markets you can walk in with cash and walk out with a full set of drone parts; Unitree does the same thing for its own motors, so they avoid the markup that rivals pay to outsource. That vertical integration lets them keep margins healthy while slashing prices, echoing the way DJI lowered controller costs or BYD drove battery prices down by making the parts themselves.

The cost drop isn’t just about selling a cheaper robot; it changes the whole experiment loop. Researchers and hobbyists can now buy a unit that will break after a few dozen runs, but they get it cheap enough to keep iterating fast. That cheap‑and‑breakable model fuels the software side too—people can test new controllers and AI models without worrying about a $30‑k price tag, and the hardware can be swapped out as the software improves.

All of this means the robot is still clunky—knees bend oddly, the walk is uneven—but the real surprise is how the manufacturing trick makes the idea of a household helper feel less like a distant dream. If you can get a machine that can fold a shirt or carry a tray for a few hundred dollars, the gap between sci‑fi and everyday chores starts to shrink, even if the robot itself still looks a bit like a toddler on stilts.

I’ve been chewing on a couple of things that just flipped under the hood this week. First off, OpenAI’s “Jalapeño” chip is already looking like a mini‑TPU: a near‑reticle die packed with 216 GB of HBM3E, bandwidth pushing 7 TB/s, and around ten petaflops of FP4 compute. What’s wild is how fast they got it from design to tape‑out—nine months, which is lightning for a high‑performance ASIC. The chip is meant to own the whole stack for LLM inference, so the economics and latency start looking a lot less dependent on the usual GPU suppliers.

On the software side, Matei Zaharia’s new Omnigent project is trying to stitch together any coding or knowledge‑work agent into a single, pluggable framework. It’s not a brand‑new miracle recipe, but the architecture is clean enough that a thousand AI‑native shops are already re‑discovering it on their own. The idea is to give you a standardized, secure way to pull in agents without having to reinvent the wiring each time.

Meanwhile, Anthropic’s Claude is moving from a neat Slack bot to a full‑blown org‑level harness. They’ve given the agent its own credentials, audit trails, and a revocable identity, which solves a lot of the “who did what” headaches but opens up new questions about scaling permissions and lock‑in. A bunch of teams are answering that by self‑hosting their own Slack‑style agents—Hugging Face’s Moon Bot, for example—so they keep the memory layer in house and avoid vendor‑specific quirks.

Finally, the memory conversation is finally getting the attention it deserves. Projects like Weaviate’s Engram and LangSmith’s “sleep‑time compute” are treating memory as a separate data‑management layer: deduplication, scoping, and lifecycle handling instead of just dumping everything into the prompt. At the same time, Chinese open models such as GLM‑5.2 keep closing the gap, showing strong results across coding

Greenspan would have looked at the AI surge the way he saw the late‑90s tech wave: a shift in the underlying cost structure. He’d point out that the real engine isn’t the hype around models, but the plummeting price of compute and the explosion of data pipelines that let firms automate tasks that used to need whole departments. That cheap, scalable infrastructure rewires incentives for capital allocation, nudging investors toward firms that can embed AI into their core processes rather than just add a flashy feature.

He’d also stress the feedback loop between policy and innovation. When the Fed kept rates low, it freed up capital for risk‑taking, and that same liquidity fed venture funds that bet on AI‑first startups. The result is a tighter coupling of monetary conditions and the speed at which new tech diffuses through the economy.

Finally, Greenspan would warn that the boom’s sustainability hinges on productivity gains translating into real output. If AI merely reshuffles existing work without adding measurable value, the surge could fizzle. The key, in his view, is whether the technology lifts the long‑run growth trend or just creates a temporary lift in asset prices.

The piece isn’t about being polite to the bots; it’s about where the math puts us. AI has gotten really good at the “mean” – the middle of the bell curve where the standard product‑manager tasks live: writing user stories, sketching roadmaps, running prioritisation grids. Those are now things a model can churn out in seconds, so competing on them is like trying to be the fastest calculator. The real edge, the article says, is in the tails.

On the left side of the curve are deep, niche expertise that only a handful of people actually own. Think of a PM who knows the gritty edge cases of cross‑border SEPA payments, the vendor quirks, the compliance gray zones that never make it into a textbook. That kind of contextual knowledge isn’t something a general‑purpose model can reliably reproduce, so it stays valuable.

On the right side are the relational, human skills that no algorithm can fake. It’s the ability to read a room, sense a CTO’s hesitation, or feel the tension before a meeting even starts. It’s the intuition about what “good” feels like for users before the data catches up, and the knack for navigating politics, personalities, and hidden incentives. Those are the parts of the job that keep a product team moving forward when the facts are fuzzy.

The takeaway is simple: stop widening your skill set and start deepening it. Use AI as a research assistant, not a replacement, and focus on building the niche knowledge and relational judgment that sit in the tails. Those are the spaces where AI still can’t tread, and they’ll keep you irreplaceable for the foreseeable future.

The thing that’s shifting under the hood is how fragile the whole revenue model is. Anthropic and OpenAI look huge on paper because they’re counting monthly run‑rate as if it were a guaranteed year‑long stream, but a single big client pulling the plug can collapse that illusion instantly. The API contracts are so easy to cut that the “steady income” they brag about is more like a honeymoon phase than a marriage.

What’s happening on the ground is a wave of cost‑cutting. Companies from Microsoft to Uber are slashing token budgets, and even internal AI‑heavy teams are being told to stop using tools unless there’s a clear purpose. The buzz that made AI feel like a must‑have is now colliding with the reality that, for many real tasks, the models still make trivial mistakes and need constant supervision.

Because of that, the whole ecosystem is stuck between “this is cool” and “how much are we paying?” – and the balance is tipping toward the latter. The math looks impressive, but it’s built on shaky assumptions: high‑volume API usage that can disappear overnight, and a lot of revenue coming from other AI startups that are themselves chasing a bubble.

Bottom line: the industry’s growth hinges on turning that flashy run‑rate into something that actually delivers value day after day. If the models don’t start working reliably in the everyday valleys where most users live, the revenue will dry up just as fast as it surged.