cvwright

You should try writing a design doc with the AI before you have it write the code. It will make random extrapolations from your first prompt, some good and some bad. Then you get a chance to argue back and forth with yourself with the robot as a helper.

The best early example of this is that Anthropic is already eating the lunch of all the new “AI security audit” companies. And they were only a few years old.

Those guys certainly thought they were being novel and creative, using AI to disrupt an expensive and labor intensive business model. But now with Claude Security, their own market share is going to be gobbled up before they can even get established.

Unfortunately even in the old days, a truly good programming book like you’re describing was depressingly rare.

Younger me really enjoyed some of the game programming books by Andre Lamothe.

Most “Learn Language X” books were terrible with over focus on syntax and very little thought into organization.

LLMs are bringing us back to all the “proper” software engineering stuff that we’ve always known we should be doing, but until now we never had enough time/people/money to do it right.

Brainstorming and research before writing a design.

Writing a design or spec before writing the code.

Comprehensive unit tests.

Etc etc etc.

Like you, I get vastly better output from the tool when I create a detailed spec in markdown before I let it start coding. And bonus, the LLM is pretty good at helping with the spec too.

You’re arguing that we have too many PhD students in CS, not too few.

I agree with you fwiw.

re: incentives, my proposal was always to let schools pay their football and basketball players, but require that grad research assistants are paid the same.

I worked a ton in grad school, and it definitely sucked at the time.

But it’s crazy to complain about getting paid to go to school. A grad stipend is there to minimally support you so you don’t have to get another job and can focus on your research. It’s not supposed to be a career!

Even that press release never claimed that Mythos was better than Opus at finding bugs.

They claim the huge advance is in exploiting the bugs.

The word “exploit” may be doing a lot of work here. In my experience Opus 4.6 is perfectly happy to provide test cases that trigger ASAN, even without the super secret squirrel security access.

But if you ask it to get you a shell it’ll probably tell you to get lost.

It’s been said that the British executed about 1% of their population each year for a few hundred years, and that a similar number died in prison.

The claim is that this made Britain a much safer country in later centuries.

New section, like pre-crime but for history. Pre-history.

And you don’t have to worry that some random passenger will piss, puke, or shit in the Waymo during your commute.

The first two happened to me within the span of a month during the three years that I rode Trimet in Portland.

Unless the data is a lagging indicator

Are Mississippi and Louisiana at the top of the pay scale?

Then why are their reading scores improving so dramatically compared to wealthier states? Especially for under-privileged populations?

Because the Brits will never be accepted by the locals as a Thai or Vietnamese or whatever, and they know it.

Whereas in western countries we have “new Germans” or “new British” from all over the world.

It’s easy to find sketchy lines of code in any large C project.

The big advance that they are claiming with Mythos is the ability to triage all the hundreds of candidate vulns and automatically generate exploits to prove that the real ones are real. And if they’re really finding 27-yr-old 0-days in OpenBSD, then it’s not just hype.

Right. It’s things like Baltimore (when I lived there) requiring that high speed internet had to roll out in poor areas first, before it could go into the rich neighborhoods.

But this was the early 2000s and the internet was still “new”. Only the richer areas cared and were willing to pay the price. Letting them have first (or even equal!) access would have made it easier to fund the rollout in low income areas.

I thought that was kind of how the hard sciences work already?

My grad school friend who was a physicist would write his talk just before his conferences, and then submit the paper later. My experience in CS was totally backwards from that.

Find-then-patch only works if you can fix the bugs quicker than you’re creating new ones.

Some orgs will be able to do this, some won’t.

You’re conveniently ignoring the Olympic boxing champion from 2024 who beat the absolute shit out of the female competitors.

Or maybe this is like fondly remembering the busted economy car that you drove around with your friends? I have my first 386DX sitting on my desk right now and it looks exactly like the top left of that photo.

The hot car that we all lusted after was maybe something like a SGI Indy or an O2.

You’re not wrong, but you probably could have built the thing with Claude in the time it took you to write this comment.

The closest thing we have is, in security / privacy / cryptography, you can write "attack" papers.

It's not perfect. You don't get any credit unless you can demonstrate a substantial break of the prior work. But it's better than in a lot of other fields.

Several people have tried over the years. We all failed, because it doesn’t work.

The economics don’t work because no one is willing to pay.

The network effect doesn’t exist, because real people don’t post enough to get the flywheel started.

All the dark patterns exist because that is what users reward.

Sucks but it’s true.

Very likely they got the causality backwards. Every time there’s a big war, technology advances because governments pour resources into it.

Just burn the tokens. It’s an upfront cost that you pay once at the beginning of a project, or on a smaller scale at the beginning of a major feature.

For context, I’ve built about 15k loc since Christmas on the $20 plan, plus $18 of extra usage. Since this is a side project, I only hit the limits once or twice per week.

It’s really not. For anything substantial, the things that you do to manage an LLM are the same things that you should be doing to manage a team of human devs, even if the team is just yourself.

Documentation. Comments. Writing a plan and/or a spec before you begin coding. Being smart with git commits and branches.

Having someone else pay for your national defense for 80 years sure doesn’t hurt.

What an understatement :)

It’s a lovely language but the compiler has got to be the most unreliable I’ve ever seen.

It crashes semi-frequently. And it will sometimes try to run analyses that are way beyond O(n). So you can have perfectly valid code that it can’t compile, until you simplify or reduce the size of some function.

They made a couple of Intel boxes in the very late 90s / very early 00s, but the company was already on the way out by that point.