Ben Thomasson

Building - persistent, justified, shareable knowledge for LLM agents.

[Ben Thomasson @benthomasson]

Take advantage of this. Put Claude in your SDLC loop. Have Claude use your software. Ask Claude what would make your software easier to use. Iterate faster than you ever had before. You can automate user testing now.

[Ben Thomasson @benthomasson]

Claude is your user. Not figuratively or metaphorically, literally. If you are writing software that has a UI, API, or library interface then Claude is your user. Humans are interacting with Claude and Claude is interacting with your software.

[Ben Thomasson @benthomasson]

From idea to release in one day: https://t.co/aj2HQe0FJH with Claude Code.

[Ben Thomasson @benthomasson]

Python is a perfect good specification language for LLMs to to produce Rust programs nearly automatically. Python allows to prototype quickly and after the behavior is defined, LLMs can tackle the type safety in Rust.

[Ben Thomasson @benthomasson]

Prototype in Python, release in Rust.

[Ben Thomasson @benthomasson]

Read more here https://t.co/AawqFNDJGC

[Ben Thomasson @benthomasson]

Keep a diary. The simplest thing that you can do to improve your effectiveness when using using AI for any task is to have the AI agent write out diary entries. Make sure the diary entries contain a date stamp or are organized in a dated directory structure.

[Ben Thomasson @benthomasson]

This is the simplest form of extended memory that you can implement yourself with any AI agent and it works very well.

[Jonathan Ross @JonathanRoss321]

For 50 years, software engineering ran on code rationing. Writing code was expensive, so we rationed it carefully through roadmaps, RFCs, prioritization meetings, and scope reviews.

This created a role: the No Engineer. No, that won't scale. No, we don't have bandwidth. No, that's out of scope. No, we need a design doc first. The No Engineer was valuable for 50 years. Every "no" saved real money. Their judgment was the rationing system.

LLMs will be the end of code rationing. Code is cheap now. And while the No Engineer is explaining why something can't be done, the Yes Engineer has already shipped three versions of it.

If you're a Yes Engineer, the next decade is yours.

[Ben Thomasson @benthomasson]

@NoahKingJr You are a bit late by 2 years and 2 months.

[Eric S. Raymond @esrtweet]

Fast, cheap AI-assisted decompilation of binary code is here. Which means code secrecy is dead.

Decompilers in themselves are not a new technology. Security researchers have employed them for years to analyze compiled malware. There's been some limited use by others, notably by hobbyists decompiling abandonware games. But there were a couple of issues that prevented this from becoming common practice.

One is simply that running decompilers was difficult. It wasn't as simple as feed in binary, get out source; it needed a person with specialist skills prepared to do spelunking through wildernesses of machine code and object formats. The other problem was that decompilation didn't give you anything like the explanatory comments that had been in the original code, so you could easily wind up with code that you could read without being able to understand or modify it.

Now large language models are busily smashing both of those barriers flat. They're better at the kind of detail analysis required to run the human side of a decompilation than humans are. More importantly, in the process of decompiling code, they rather automatically build a global model of how it works that can easily be expressed by high quality comments in the extracted code. All you have to do, basically, is ask for the comments.

I'm going to reinforce that latter point because it may not be obvious how good LLMs are at this, and how much better they're going to get. When they decompile code and comment it for you, they're not just working from that one piece of code you have put in front of them - they'll have in their training set hundreds, possibly thousands of pieces of code similar to it and with comments. This will give them superhuman levels of insight not just into what it does at the microlevel, but what it means to the humans who wrote it, and what technical assumptions it's embodying.

Compilation no longer guards your secrets. Or, to put it more precisely the expected time span in which you can still count on it to obscure them is measured in months. Possibly weeks.

What does this mean?

It means you're in an open-source world now. All it's going to take for anybody to bust your proprietary IP open is care enough to spend tokens on the analysis.

You will maximize your chances of survival as a software business if you get out ahead of this rather than trying to fight it.

This isn't exactly the way I expected open source to win. But, you know, I'll take it. Good enough.

[Ben Thomasson @benthomasson]

@plainionist Python is the pseudo code spec for your agent to write the right rust code. Agents think in code probably better than natural language.

[Ben Thomasson @benthomasson]

@gvanrossum Cheating is a valid heuristic for AI. Multiple adversarial agents with different contexts and goals are needed. One agent is not enough

[Seb @plainionist]

Hot take:

Rust might be the best language for vibe-coding.

The compiler provides strong feedback
and guardrails.

And AI thrives on guardrails.

[Ryan Hart @thisdudelikesAI]

This isn't a theoretical optimization.

If you run latency-sensitive workloads, HFT infrastructure, real-time systems, game servers, or anything where p99 latency actually matters, DRAM refresh stalls are taxing you on every machine right now.

Tailslayer is the first open source library that actually fixes it.

Repo: https://t.co/JgcK2vaRhi

100% Opensource. Apache 2.0 License.

[Ben Thomasson @benthomasson]

@mathelirium Wings throw air down due to the boundary layer effects pulling air mass above the wing down after the wing passes it. Down draft is real.

[Ben Thomasson @benthomasson]

@rohit4verse No we don’t. AI can do the whole SDLC loop every 15 minutes.

[Ben Thomasson @benthomasson]

@vitrupo AI can write the spec too. Show it the problem.