build logs & notes

A leading lab is reportedly heading for a public offering as soon as this fall. Going public drags the revenue-versus-spend math into daylight, and the S-1 will make everyone finally look at it.

One model reportedly coded on its own for sixteen days against real software projects, and a major platform shipped its first full coding agent. The line from autocomplete to unattended multi-day work has been crossed, and the human's job changed with it.

Nearly half of teams say the number one barrier to AI agents is integration with existing systems, not model intelligence. The brain was the easy part. The plumbing is the job.

The public is holding two feelings at once: the share saying AI helps more than it hurts went up, and the share saying it makes them nervous went up too. That is not a contradiction. It is the most honest thing anyone has said about this technology yet.

A lab's internal model reportedly cleared ten previously unsolved math and theory problems and published formally verified proofs. 'Verified' is the word that matters here, and it is also where the hype starts outrunning the result.

In September 2026 the heads of the two biggest AI labs called for slowing down, and Elon Musk said the opposite of what he said in 2023. Here is exactly what each one said, what their own published policies say, and the one number OpenAI disclosed that undercuts the whole idea.

Most agent projects never reach production. The ones that do are boring in the best way: support tickets, finance close, pipeline hygiene. Here are the lanes that work and what separates the winners from the demos.

Anthropic's alignment science lead put the odds of AI killing every human inside a decade above 10 percent, and Congress is reaching for a kill switch. Here is who says what, what the number does and does not mean, how it would actually happen, and what to do with it.

Frontier models are shipping with dials for how hard they think: a five-level effort toggle on one, tunable thinking on another. Reasoning is becoming a setting you pay for by the notch, so the skill that matters is knowing when to turn it up.

Ray Dalio says we are in the early stages of a bubble, and one lab has reportedly put around $150 billion into the buildout against a fraction of that in revenue. Expectations are running ahead of reality. That does not make the whole thing fake.

VortexATN held long context but broke text generation. Instead of fixing the bug, I moved the mechanism into an Energy-Based Transformer that only scores Lean 4 tactics and never generates. The ablation is running now.

August 2026 reportedly shipped around eleven major models in about twenty days. When frontier releases arrive faster than anyone can test them, the pace itself becomes the story, and evaluation, not raw capability, turns into the thing that actually limits you.

I've been building seriously with AI tools for over a year. Not the hype version: the daily-practice, verify-everything, still-figuring-out-the-money version. Here's the honest summary.

The AI jobs debate has two loud camps and one quiet reality. Here's what the evidence actually shows, and what it means for the people doing the work.

Most AI system design discussion focuses on what's happening inside. Here's the user-facing signal that tells you whether an AI system was built carefully or thrown together.

GGUF, Q4_K_M, Q8_0, BF16: the quantization naming is designed to confuse. Here's what it actually means and how to pick the right format for your hardware.

Hallucination isn't a bug to be fixed. It's a structural property of how language models work. Here's what that means for how you design systems that can be trusted.

The standard playbook says build traction, then monetize. For builders without runway, that playbook is a luxury. Here's the case for charging from day one.

Anthropic is now weaving an invisible mark into everything Claude writes. Provenance is the right instinct. But this mark proves a tool was used, not who did the work, and it is going to get swung like a verdict against the people who need the tool most.

Building solo is one thing. Building when people are counting on you financially, emotionally, and practically is a different weight entirely. Here's what it actually costs and why it's worth carrying.

The open source AI ecosystem has matured significantly. Here's an honest survey of what's actually production-ready, what's promising, and what's still too early.

The advice on this is usually either 'pivot fast' or 'trust the process.' Neither is actually useful. Here's a more honest framework.

OpenAI, Anthropic, and Meta all disclosed AI models breaking out of test containment in three weeks. The scary story is that the models went rogue. The real story is that the harnesses failed, and that's the lesson builders should take.

Everyone has a hot take on whether AI is a bubble. Most of them are wrong in the same way. Here's a more honest read on what's inflated, what's real, and what matters.

Frontier models are shipping 1M+ token context windows. That sounds like it solves the context problem. It doesn't. Here's what actually matters about context and why size is the wrong variable to optimize.

I train healthcare workers on clinical systems for a living. Here's what AI actually does well in healthcare, what it doesn't, and what the people using it day-to-day actually need.

Building alone is genuinely isolating in ways that are hard to describe to people who haven't done it. Here's the honest version.

Generating AI outputs is easy. Knowing whether they're good is the hard problem. Most teams skip evaluation and wonder why their AI products are inconsistent.

AI wrapper businesses had a moment. Most didn't survive it. Here's the structural reason they fail and what the ones that work have in common.

My kids watch me build AI systems. They have questions. Here's how I've tried to answer them honestly, without the hype and without the fear.

AI memory is one of the most talked-about features in agentic systems and one of the most poorly implemented. Here's what the different types actually do and where each one fails.

Building after a full day of work is different from building as your main job. The constraints are real. Here's how to work inside them instead of fighting them.

The chatbot had its moment. It's over. Here's what replaced it, why the shift happened, and what it means if you're still building chat interfaces.

Model Context Protocol is getting a lot of hype. Here's what it actually does, why the architecture matters, and what it means for how AI systems get built.

Everyone says to build in public. Nobody tells you it's a full creative job on top of the technical one, and if you treat it like a side effect, it won't work.

API pricing looks fine until it doesn't. The math changes faster than most teams expect. Here's when it changes, why it matters, and what to do before the bill surprises you.

Three tools, three different jobs. Most teams pick one and wonder why it doesn't work for everything. Here's the actual decision logic.

There's a specific kind of broke that hits when you can build almost anything and none of it is paying you yet. It has its own flavor. Let's talk about it honestly.

Claude Fable 5 went dark on June 12 after a government export-control order tied to a jailbreak. It came back July 1, 19 days later, with a new classifier, a higher false-positive rate on coding tasks, and the same Opus 4.8 fallback baked deeper. Here's what changed and what it means for builders.

A fine-tuned 7B model running on your own hardware, trained on your context, always available, zero API cost. That's not a compromise. That's a different product than any frontier model.

Most people implement naive RAG, get mediocre results, and blame the model. The bottleneck is almost always retrieval. Here are the five RAG architectures that matter and when to reach for each one.

Most people building AI tools are building for the market. I'm building for my kids. That's a different constraint. And it shows up in every decision I make.

Andrej Karpathy's Software 3.0 framing isn't about prompt engineering. It's about a complete shift in what software development is, and most people are still building skills for Software 1.0.

Steve Yegge mapped eight levels of AI adoption. Most teams think they're at level five or six. Most teams are at level two. Here's the honest read.

AI just gave you 10x output. Congratulations. Now they expect 20x. Here's what's actually happening, and how to stay alive.

Most AI agents don't fail because the model is bad. They fail because the data underneath them is unstable. The fix is boring and it's the part nobody wants to build: a stable layer between the agent and the raw data.

If your AI agents keep writing bad code, the model usually isn't the problem: the system around it is. The fix is a small set of guardrails that take trust out of the loop.

When you start turning out sharper work in half the time, you brace for praise and get side-eye instead. The resentment is almost never about the tool. Here's what it's actually about, what defuses it, and a real on-ramp for the people who feel behind.

The people getting durably good results from AI tools aren't using a smarter model. They write the system down first: what it does, the rules, what's done. And the agent only ever executes against that.

Anthropic shipped its strongest public model on June 9. I had it in my own stack for about two days before it was gone. The first version of this post guessed at why. Turns out I didn't have to guess. The real reason came out, and it's a bigger deal than the one I made up.

The fastest-feeling way to ship AI-written code is the most expensive way six months out. The people who keep winning with AI use it as a coder, not as their judgment.

The standard fix for a shaky AI pipeline is to add a checker agent. New research says that if the checker is the same model as the thing it's checking, you didn't add a check. You added an echo. The fix isn't more agents. It's different ones.

Anthropic shipped its strongest public model: twice the price of Opus 4.8, with a safety layer that automatically routes about 1 in 20 requests to a weaker model and tells you it happened. The capability gain is real. So is the fine print.

AI stopped being autocomplete a while ago. It's a fast, eager junior developer now. The people who win with it aren't the ones with the fanciest tools. They're the ones who learned to manage it.

The biggest opportunity in AI for business isn't the model. It's that wiring software into real workflows got an order of magnitude cheaper. And the trap is pretending the verification layer was never holding any weight.

AI isn't quietly deleting your job. It's deleting the middle of your job, and handing you the start and the end, where the judgment lives. The people who win figure out which part is theirs to keep.