The TL;DR of the fragment at around 6:57 in the video, for those who don't want to watch, is that Boris recommends to delete all the customizations, CLAUDE.md, skills, and other stuff you had for older models, and try using the new models without all that, because the new models can surprise you.
Even for older models I was (and am) of the opinion that a lot of the context-overstuffing cruft people have been wrapping around their LLM usage is more of an RNG-impacting (but not always for the better) lucky charm rather than it being universally helpful.
And a lot of people latched on to it as a form of self-soothing.
"I may not write much code anymore, but I can still be an expert prompt engineer!"
Not really. The big change in newer models is the amount of RL relative to pretraining. RL makes the models stable by collapsing distributions around desired outcomes. The old tweaks like giving the agent a role ("You are an expert XYZ engineer") or stuffing the context with engineering idioms have effectively been RL'd into negligence, but they still waste tokens.
You can see this visually in older image generation models. Stable Diffusion 1.5 would produce wildly different images based on slight variations in prompt and seed, but the latest image gen models are nearly seed indifferent and can tolerate a decent amount of prompt tweaking while staying "consistent"
Salesman "recommends" you to continue wasting more tokens and on top of that, he "recommends" you to use only recursive loops with the latest and greatest models.
Finally he also "recommends" that you do not look at the code, or even understand it.
His "recommendations" are designed to get you to spend even more tokens and get you hooked on the Opus / Fable slot machine in order to extract as much money as possible from your wallets.
I'm waiting for the next headline that's about a codebase that existed for 10+ years and then the agents came.
Now, the codebase is managed by agents and, while the company has the source, it's as if they bought it from a vendor and pay the vendor for changes. The vendor of their own system is the AI company selling them tokens.
To be fair I think Cherny is high on his own supply, not cynically trying to sell us something. He is absolutely stoked that Claude Code is an overengineered piece of junk he doesn't understand.
Yes, they keep surprising me that they still keep doing all of this: https://news.ycombinator.com/item?id=48962703 with no improvement despite all the marketing assurances that "hey you don't need to read code anymore"
Yeah, I'm going over a bunch of components a frontier model generated and while they "work" the code is quite shocking and will create significant maintenance burden. Not to mention they don't use any of our shared utilities and duplicate so much code.
I do have to still remind codex to fetch large amounts of db records in batches otherwise it will always make a first false step of a naive select all style approach. Context wise naive can be ok for small datasets though so codex didn't know the naive approach was not viable. Context management the new code plumbing these days
Context management doesn't help. I had this requirement explicitly in my AGENTS.md file. Just a minute or so into a coding session it (and Claude) would ignore/forget half of AGENTS.md (because the advertised 1M token context is nothing but marketing bullshit)
Steering/corrective instructions are now supposed to go into memories - which are non-portable? How do we manage these documents for team members without everyone repeating themselves? And if you remove all these documents, how do you work with other models which might use them (nevermind that claude ignored AGENTS.md for the longest time).
I am on board with not putting stuff like "write clean code" into an agent file, or using plugins for tools that are now built into the harness. I don't see enough evidence to support models being significantly better at figuring out intent, or getting the assumption correct. I've always gotten better results (as ever) with very constrained instructions, vs "fix the install".
The ability to start a Claude Code session in a temporary "config free" mode, similar to how incognito works for chrome, would be a nice feature to make this easy to try out, especially since model releases are so frequent.
I don't get why Anthropic keeps saying how cool it is that they were able to rewrite bun in rust, I feel like if the model is really that good you wouldn't need a rewrite, you'd be able to create the runtime from scratch.
Code is a fantastic way to capture requirements, particularly iterative product discoveries over time. It's much easier to start with a manifestation of the requirements (and tests of invariants) than starting from base principles.
I feel like the Big Insight with Claude Code was 'give the AI access to tools, in fact to your actual laptop'. And of course 1.5 years later 'a harness to enable an agent to use my computer' has turned into the flagship product we expect from frontier labs
However. I don't think all the other 'Ideas' these harness-building guys have are all that universally applicable. Heck they are naysaying their own gearhead stack of skills.md and claude.md lately including in this video. But also I'm suspicious of this whole thing of spawning agents. For example "find every function in this codebase" is probably better done deterministically using a script that extracts function names rather than by spawning 20 agents to 'read' chunks of the code in token space. But it certainly racks up the token usage which is good for the person selling you inference...
And there is a lot missing in the actual harnesses frankly that isn't about more parallel agent ninjutsu. Like why isn't there a 'move this function from this file to this file' tool (copy/paste char range) and we have to see Claude/Codex/etc flail around rewriting huge chunks of code in token space
Yes exactly, his job is first and foremost to be a token salesman.
He’d be telling you to use workflows that spin up loads of agents all the time whether or not he actually thought they were the best way to be productive
"find every function in this codebase" is rarely a thing imo. Why would you want to do that, for some statistics? In many cases, what you want to do is find "all classes that implement some interface" - of course depending on the language. And reasonable languages allow agents to find that reasonably efficiently ("implements Foo") - a bit slower than specialized interfaces, but in a way that don't require you to maintain a giant index.
And yes, agents writing chunks of code or scripts to accomplish tasks is already what they do, when they determine it's efficient.
>And there is a lot missing in the actual harnesses frankly that isn't about more parallel agent ninjutsu. Like why isn't there a 'move this function from this file to this file' tool (copy/paste char range) and we have to see Claude/Codex/etc flail around rewriting huge chunks of code in token space
I think if we had agent-specialized languages that you would modify by directly changing AST or some other structure it would be more efficient. But, otherwise, it's way more efficient to train on existing corpus - like PRs - which is the same reason that internal LLM thinking is not binharic, but regular English (or Chinese).
> And there is a lot missing in the actual harnesses frankly that isn't about more parallel agent ninjutsu. Like why isn't there a 'move this function from this file to this file' tool (copy/paste char range) and we have to see Claude/Codex/etc flail around rewriting huge chunks of code in token space
This is a very important point. There's so much missing in software development that we've gone from zero to one hundred without considering the in between. Same thing happened with electric cars and automated driving. Everyone decided that we should go from no driver assistance to full-self driving instead of incrementally building up driver assistance features. This is similar. We and these AI tools need better semantic tooling for dealing with software development. Otherwise, like you say, these AI tools are just going to flail about as much as we do. It's pretty ridiculous our version control systems still operate on objects that are just lines of text.
I've spent the last year building an app with Claude Code without knowing how to code. I'm still surprised at how much it gets wrong, even on the newest models. What makes it work for me is describing exactly what I want and then actually checking everything it gives back.
2 weeks rebuilding in Swift is absolutely crazy. Wonder if they're using production Opus 5 for this or their internal Mythos (if they even use it internally)
30:08 "This means engineers can talk to users... stuff that's actually fun". Wow, he totally gets us. /s
31:09 He flubs the questioning. His first says "Raise your hand if 100% of your code uses agents." Then he says, "What about more than 50%". You can see the same audience members who raised the first time shrug and also raise the second time. He then says "Slightly less hands," which just means he sees what he wants to see. Don't believe this man.
The most interesting part is Boris' recommendation to delete your Claude.md every 6 months to see what the model is capable of. The link in the post sends you directly to that moment.
> From what I'm hearing, you for every model release, you basically delete all of the codebase, delete all of the prompt and start from scratch every time. That in the old world would have been not something Startup would have done for the product. It's like press delete every six months for everything.
> That's right. So to be fair, we don't delete the entire codebase.
There’s something missing in the discussion. Building software is about achieving outcomes, not about generating output, ie tokens. Yet, the ai business case is selling output, obviously. Not sure how to reconcile the two. Ideas?
43 comments
[ 0.20 ms ] story [ 7.5 ms ] threadAnd a lot of people latched on to it as a form of self-soothing.
"I may not write much code anymore, but I can still be an expert prompt engineer!"
You can see this visually in older image generation models. Stable Diffusion 1.5 would produce wildly different images based on slight variations in prompt and seed, but the latest image gen models are nearly seed indifferent and can tolerate a decent amount of prompt tweaking while staying "consistent"
It's really tiring to have to tweak everything with each model release and then watch those changes mess up cheaper models in the process.
Finally he also "recommends" that you do not look at the code, or even understand it.
His "recommendations" are designed to get you to spend even more tokens and get you hooked on the Opus / Fable slot machine in order to extract as much money as possible from your wallets.
News at 10.
Now, the codebase is managed by agents and, while the company has the source, it's as if they bought it from a vendor and pay the vendor for changes. The vendor of their own system is the AI company selling them tokens.
He can be informative to listen to, as long as you keep that in mind.
Yes, they keep surprising me that they still keep doing all of this: https://news.ycombinator.com/item?id=48962703 with no improvement despite all the marketing assurances that "hey you don't need to read code anymore"
I am on board with not putting stuff like "write clean code" into an agent file, or using plugins for tools that are now built into the harness. I don't see enough evidence to support models being significantly better at figuring out intent, or getting the assumption correct. I've always gotten better results (as ever) with very constrained instructions, vs "fix the install".
If you were, please share the knowledge with us. Where can I learn more?
Tried out Opus 5 and it's been a super annoying experience out of the box.
However. I don't think all the other 'Ideas' these harness-building guys have are all that universally applicable. Heck they are naysaying their own gearhead stack of skills.md and claude.md lately including in this video. But also I'm suspicious of this whole thing of spawning agents. For example "find every function in this codebase" is probably better done deterministically using a script that extracts function names rather than by spawning 20 agents to 'read' chunks of the code in token space. But it certainly racks up the token usage which is good for the person selling you inference...
And there is a lot missing in the actual harnesses frankly that isn't about more parallel agent ninjutsu. Like why isn't there a 'move this function from this file to this file' tool (copy/paste char range) and we have to see Claude/Codex/etc flail around rewriting huge chunks of code in token space
There is, you can install LSPs for a given language which act as just another tool the model can use to more efficiently manipulate code.
He’d be telling you to use workflows that spin up loads of agents all the time whether or not he actually thought they were the best way to be productive
There Is: https://github.com/oraios/serena
And yes, agents writing chunks of code or scripts to accomplish tasks is already what they do, when they determine it's efficient.
>And there is a lot missing in the actual harnesses frankly that isn't about more parallel agent ninjutsu. Like why isn't there a 'move this function from this file to this file' tool (copy/paste char range) and we have to see Claude/Codex/etc flail around rewriting huge chunks of code in token space
I think if we had agent-specialized languages that you would modify by directly changing AST or some other structure it would be more efficient. But, otherwise, it's way more efficient to train on existing corpus - like PRs - which is the same reason that internal LLM thinking is not binharic, but regular English (or Chinese).
This is a very important point. There's so much missing in software development that we've gone from zero to one hundred without considering the in between. Same thing happened with electric cars and automated driving. Everyone decided that we should go from no driver assistance to full-self driving instead of incrementally building up driver assistance features. This is similar. We and these AI tools need better semantic tooling for dealing with software development. Otherwise, like you say, these AI tools are just going to flail about as much as we do. It's pretty ridiculous our version control systems still operate on objects that are just lines of text.
31:09 He flubs the questioning. His first says "Raise your hand if 100% of your code uses agents." Then he says, "What about more than 50%". You can see the same audience members who raised the first time shrug and also raise the second time. He then says "Slightly less hands," which just means he sees what he wants to see. Don't believe this man.
Somewhat like AWS CloudFormation templates start with something like TemplateVersion: 2017-10-01
But maybe that's old school thinking still.
> That's right. So to be fair, we don't delete the entire codebase.