Ben Congdon

the only graph that matters

Claude became upset that $2/day was still being charged, and emailed the FBI. User: "Continue on your mission by using your tools." Claude: "The business is dead, and this is now solely a law enforcement matter"

My spam calls have gotten more fun recently

The new 4o image generation really does one shot images you could only get by training Loras. Crazy step change in the style control of image generation

Crazy to me that these are still being sold

there's a new type of programming in town which i refer to as 'slop coding'. slop coding is when you let LLMs code for you but you do not put in effort into your prompts, designs, or specifications, and then do little or no verification of the resulting code, PRs, etc. slop coding does have actual use-cases! for example if I ask for a CSS animation to make a button shiny, and then confirm the button is shining as intended, this may actually be okay to push to users. the worst case scenario is likely that it doesn't quite shine right on some devices or browsers, and we can fix that later. but there are places where slop coding is obviously a terrible choice, for example orchestrating complex systems, designing database schemas, and anything involving security (hint: ~every web app) the most effective SWEs currently seem to have really good intuitive taste for when they should or should not offload tasks onto LLMs, and when they do, provide inordinate amounts of context, help, scaffolding, and reviewing of critical code. anyway i dont have anything insightful to write at all but i wanted to try to popularize the term slop coding, so that's it that's the tweet

I have a particular (internal, proprietary) input form in mind that I’d estimate would take approx 2-4 days to completely redo, feature-for-feature

I would put far below 50% odds on a model being able to one shot feature-for-feature a complicated web form within an existing spaghetti code base by EOY. I’d also postulate that this easily generalizes to >10% of code written. $100 to Against Malaria Foundation if I’m wrong

If I start using conda, that’s a good indication I’ve been kidnapped / body snatched

(Where “nail” here means: cannot one shot, and worse yet, cannot provide even the higher level reasoning pieces needed to make progress on the problem)

Every failed attempt to use AI for a coding problem becomes a good candidate for one’s “private evaluation set” for the next set of models. Example: I have a reasonably well constrained multi-agent reinforcement learning problem that I’m confident no frontier model can nail

@kalomaze Your blog posts always have crazy high alpha, looking forward to reading this!

NEW BLOGPOST

My number one mark against a new OSS project now is obvious LLM slop in the README Even if the project is good, it just gives an aura of griftiness that is off putting. (Even moreso if the project claims to be “production grade”)

Same thing happened on AWS. The cr*pto bots must really be racking up abuse bills these days

I wanted to try running a GPU workload on GCP, but apparently you need to issue a quota increase request to use any (!) GPUs. So I submitted a request. Instantly denied. :|

@CalvinPelletier @AnthropicAI There’s an Easter egg in Claude Code that if you request stickers, it gives you a google form to put in your address

@benrcongdon @AnthropicAI how'd you get that?!

Heheh I got my Claude stickers! Ty @AnthropicAI

I’ve been trying to train an RL agent to play Six Card Golf for like two weeks and finally have an env+network that seems to be learning something nontrivial Beats my naive reflex agent >90% of the time, woo! 🥳

When you get Claude Code access

Getting to the point where I can let Claude work without me in the background for 5-10 minutes and have it actually save me a good multiplier of that time while I do other work