Issue 83 — 2026-08-14
China's DeepSeek just broke Western AI's biggest lie
DeepSeek R1 is a Chinese AI model that beat OpenAI o1 on major benchmarks while spending roughly 1/100th of what OpenAI spent training it. The company behind it, DeepSeek, claims this is a "deep thesis" about how China can challenge Western AI dominance not through spending but through smarter architecture.
The Western AI story has been built on one assumption: more money equals better results. OpenAI spent $300 million training o1 on thousands of NVIDIA chips. DeepSeek spent about $6 million and still beat it on math, coding, and reasoning benchmarks. They didn't just cut corners — they redesigned the architecture itself, using a Mixture of Experts system where the model routes different questions to specialized sub-networks rather than running one massive model on everything.
There's also the hardware angle. Western companies have been building on NVIDIA's H100 chips, which the US has been restricting for export to China. DeepSeek built its model around domestic Chinese chips instead. If China can do more with less, the whole Western AI advantage built on spending and chip access starts looking much thinner.
Why this matters for us: if AI gets cheaper and more accessible, the companies that get built will be different than the ones we've been expecting — and the people building them won't all live in Palo Alto.
Musk hints Starlink is coming to future Teslas, Cybercab shows integrated dish
Musk is hinting that Starlink satellite internet will be built into future Tesla vehicles, with the Cybercab showing an integrated dish on the roof.
The move would give Teslas a way to stay connected even in places with weak cell coverage. Starlink's satellite network already reaches remote areas where cell towers don't, and putting a built-in dish in the car changes the game for cross-country travel and rural driving.
The Cybercab's integrated dish is the first concrete look at what this could mean for Tesla's lineup going forward. The company has been pushing connectivity hard—full self-driving needs constant data, and OTA updates require a reliable connection. Starlink would solve both problems at once.
Why this matters for us: la gente en el campo, los primos que manejan de Texas a California, los que dependen de la conexión para el trabajo—esto les cambia la vida.
Nucleus sampling: how AI keeps its head straight
When you ask an AI a question, it generates a bunch of possible next words, each with a confidence score. Nucleus sampling (called "top-p" in the docs) is a trick that keeps the AI from going off the rails.
The AI gathers up its most-likely answers until they add up to some percentage — say 90%. It throws out everything below that line and picks randomly from the rest.
Think of a bodega. The owner doesn't stock every item in the world. He stocks what people in the neighborhood actually buy. The rest of the shelf stays empty. Nucleus sampling does the same thing: it keeps the answers on the shelf and clears out the junk.
Without nucleus sampling, the AI can ramble, invent, or circle back to itself. With it, responses stay focused and grounded.
Why this matters for us: AI is getting used more in schools, government services, and customer support — places where answers that sound confident but are wrong can do real harm.
When you're testing AI outputs, ask: does it repeat itself, or does it keep coming up with new, relevant points? Repetition usually means the nucleus is too narrow.
If you build it, they may not come.
— links.tldrnewsletter.com
#bluesky-opens-the-protocol-to-outside-services-33197eThe DeepSeek Thesis: How China's AI Models Are Reshaping the Arms Race
An essay from China Talk Media makes a case that has been bouncing around Silicon Valley for a while now: Chinese AI models can compete with the best US models at a fraction of the cost. DeepSeek, the model behind the news, is the poster child—able to match top-tier American AI performance while burning significantly less compute.
That matters because it upends a foundational assumption of the AI arms race. The conventional wisdom has been that training frontier models requires massive capital, massive data, and months of GPU time.
TradeWork: jobs, crews, and payments — all in one place
Your crew is on site. The estimates are in a notes app. Invoices are PDFs. Payments come by check. The foreman texts the office, the office texts the GC, and nobody has a clear picture of what's actually done.
This is what trades businesses have been stuck with. The general-purpose CRMs are built for offices. The trade apps are built for managers. Neither is for the truck.
TradeWork is a mobile-first platform for trades crews who work out of a truck. Jobs, crews, invoices, and payments in one place. Bilingual surfaces so the office, the foreman, and the helper can all read the same job.
Esto te toca.
https://tradework.work
Cuando las predicciones tech se quedan en nada
La newsletter TLDR está destacando el ensayo de Steve Yegge, "The Shape of Things to Come" (2016), escrito cuando dejó Google. El texto revisa una serie de predicciones tecnológicas de la industria — muchas de ellas firmes, algunas firmísimas — y las compara con lo que realmente pasó.
El patrón es tan viejo como el sector: cada ciclo, los grandes nombres hacen pronósticos con la convicción de quien ha visto su software multiplicarse en la vida real. Luego el tiempo pasa, y las predicciones se quedan en el aire.
Zvi Mowshowitz's latest internal updates
Zvi Mowshowitz is posting further developments on his internal newsletter, covering updates from his latest round of thinking and projects. The piece is part of his ongoing series of internal posts that have been circulating in tech and AI circles.
These updates tend to cover a mix of project work, personal reflections, and ideas that don't fit his public posts. Zvi's internal notes have become a notable thread in the broader AI and tech conversation, particularly among people who follow his work on LessWrong and his Substack.
Why this matters for us: The people writing the internal posts that shape how AI gets talked about publicly are mostly the same tech insiders — the community needs its own voice in that conversation, not just a reaction to what they're working through.
Spec-Driven Development: How AI Writes Code Better When You Write First
AI code generators are decent when you give them a clear spec — vague requirements and they hallucinate. Spec-driven development flips the AI workflow: you write a detailed description of what the code should do, then let the model generate from that. The result is code that actually matches what you need.
It's the same principle that's been around for decades, just with a different tool. You know the cousin who builds side-apps for $20k a pop — they already write specs before they touch the keyboard.
Spec-Driven Development: AI software that actually ships on spec
DerivaAI is testing a new angle on AI-powered coding: instead of letting models freestyle, you write a spec first—like a contract—then the AI generates code that matches it exactly. The idea is to sidestep the endless back-and-forth where LLMs hallucinate features nobody asked for.
TLDR Product's take is that spec-driven development could be the missing piece in AI software engineering. Right now, most developers use AI as a co-writer, which means constant revisions. If you can lock down requirements upfront, you might get fewer hallucinations and faster delivery.
Why this matters for us: la gente que trabaja con código—desarrolladores, freelancers, cualquier primo que le cobra por hacer una app—puede que se encuentre con que la IA ya no solo escribe código sino que sigue instrucciones. Si el modelo respeta el spec, eso cambia cómo cobras, cuánto tiempo te lleva el trabajo y qué tan confiable es tu producto final.
Nobody's Actually Prioritizing Value in Product
Ant Murphy at TLDR Product dug into a quiet rot in product management: teams say they prioritize value, but the data doesn't back it up. They looked at how product teams actually rank their work—most don't. Instead they ship features, chase shiny objects, and call it strategy.
The problem isn't that people don't care about value. It's that nobody has built the muscle to actually measure it, compare it, and say no to things that look good but don't move the needle. You see it in every corner of the industry. The cousin who runs his side business?
Musk says Tesla cars will get Starlink built in
Elon Musk dropped a hint at the Tesla Cybercab reveal: future Teslas will come with Starlink dishes built right into the roof. You won't need to buy a satellite internet kit for the road — it's part of the car now.
The move makes sense on paper. Tesla already uses Starlink in some of its vehicles, and the Cybercab shows the dish integrated into the body. But the real question is who gets access, and at what price. Starlink has been a lifeline for rural folks, remote workers, and communities the internet forgot. Fold it into a car, and you're pricing out the people who actually need it most.
This is the kind of play Musk's been pulling for a while — bundle everything together, lock you into an ecosystem, and let the bill come later. Tesla owners already pay for Autopilot, Full Self-Driving, and premium connectivity. Starlink on the car is just the next line item.
Why this matters for us: when satellite internet gets folded into a premium car, la gente who need reliable connectivity on the road — truckers, rural families, anyone working between towns — get priced out of something that should be a public utility.
Vercel ships a template for building your own AI software factory
Vercel Labs released an open-source template for building AI software factories—systems where AI agents handle the full loop from code review to deployment. The project lives on GitHub and is meant to be forked and adapted.
Software factories are the latest rebranding of what has been a long-running push: let AI handle the heavy lifting in software development. You give it a task, it writes code, tests it, fixes bugs, and ships. The template Vercel published is a starting point for teams that want to stand up that kind of workflow without building everything from scratch.
Why this matters for us:
OpenAI launches ultrafast API tier for GPT-5.6 at $50/min
OpenAI just dropped a new API tier for GPT-5.6 that guarantees response times under five seconds. The catch: it costs $50 per minute, and you need to contact their sales team to get access. This is aimed at high-end enterprise customers who need real-time AI for things like customer service bots, autonomous systems, and apps where every second of delay hurts the user experience.
The pricing is roughly 5x what a standard GPT-5.6 API call costs. OpenAI is essentially creating two lanes on the highway — one for people who need speed and can afford it, and one for everyone else. The ultrafast tier is part of a broader push into the enterprise market, where OpenAI competes with Anthropic, Google, and Meta for the biggest contracts.
Why this matters for us: la gente que maneja side businesses y apps no se va a poder acceder a esta velocidad por el precio, así que si estás construyendo algo que depende de AI rápida, hay que saber que va a costar el doble de lo que ya cuesta.
pgGraph lets you run graph queries directly in PostgreSQL
A new library called pgGraph lets you run graph queries directly in PostgreSQL instead of spinning up a separate graph database. It's built on top of the existing PostgreSQL ecosystem — no new infrastructure to manage.
At the same time, a small model evaluation tool called smevals is getting attention for letting teams test smaller language models without needing a massive compute budget. The idea is simple: you can compare how different models perform on specific tasks before committing to the bigger, more expensive ones.
White Fiber is betting on cross-data center networking
White Fiber, a networking company, published a blog post about cross-data center networking solutions. The piece pitches a way for companies to connect data centers across different locations so traffic can move between them instead of going through the public internet.
The broader push here is that companies are running services across multiple cloud regions and data centers, and they want faster, more predictable connections between them. That's been a real problem — routing through the public internet adds latency, and the cost can add up when you're moving terabytes of data.
Why this matters for us: The companies building this infrastructure are the ones deciding how we store data, who controls the pipes, and what it costs to actually run a service — decisions that ripple down to every small business that needs cloud storage or runs a side hustle on the internet.