LLM Year in Review

I don't even see much reason to use Cursor. I am used to IntelliJ IDEA, so I just downloaded the Claude Code plugin and basically now I use the IDE only for navigating in the code, finding references and reviewing the code. I can't even remember the last time I wrote more than 2 lines of code. Claude Code has catapulted my performance at least 5x if not more. And now that the cost of writing test is so minimal I am also able to achieve much better (and meaningful!) test coverage too. The AI agents is where the most productivity is. I just create a plan with Claude, iterate over, ask questions, then let it implement the plan, review, ask to do some adjustments. No manual writing of code at all. Zero.

Do you guys all work 100% on open source? Or are you uploading bits of your copyrighted code for future training to Anthropic? I hate patents so copyright is the only IP protection I have.

I first got into agentic properly with GLM coding plan (it's like $2/month), but I found myself very consistently asking Claude to make the code more elegant and readable. At which point I realized I was being silly and just switched to Claude code.

(GLM etc. get surprisingly close with good prompting but... $0.60/day to not worry about that is a no brainer.)

Nano Banana Pro is legitimately an insane tool if you know how to use it. I still can’t believe they released it in the wild

I've used all of these tools and for me Cursor works just as well but has tabs, easy ways to abort or edit prompts, great visual diff, etc...

Someone sell me on how Claude Code, I just don't get it.

I don’t have much time to evaluate tools every months and I have settled on Cursor. I’m curious on what I’m missing when using the same models?

I used to use a lot of em dashes normally in my writing - they were my go-to replacements for commas and semicolons

But I had to change how I write because people started calling my writing “AI generated”

Very broadly, AI sentence-structure and word choice is recursing back into society, changing how humans use language. The Economist recently had a piece on word usage of British Parliament members. They are adopting words and phrases commonly seen in AI.

We're embarking on a ginormous planetary experiment here.

You’re absolutely right!

Jk jk, now that you pointed it out I can’t unsee it.

I hated these sentences way before LLMs, at least in the context of an explanation.

> it's not just a website you go like Google, it's a little spirit/ghost that "lives" on your computer

This type of sentence, I call rhetorical fat. Get rid of this fat and you obtain a boring sentence that repeats what has been said in the previous one.

Not all rhetorical fats are equal, and I must admit I find myself eyerolling on the "little spirit" part more than about the fatness.

I understand the author wants to decorate things and emphasize key elements, and the hate I feel is only caused by the incompatible projection of my ideals to a text that doesn't belong to me.

> it's not just about the image generation itself, it's about the joint capability coming from text generation.

That's unjustified conceptual stress.

That could be a legitimate answer to a question ("No, no, it's not just about that, it's more about this"), but it's a text. Maybe the text wants you to be focused, maybe the text wants to hype you; this is the shape of the hype without the hype.

"I find image generation is cooler when paired with text generation."

Yeah, came to read Karpathy's thoughts, but might as well ask an LLM myself..

I cannot unsee this anymore and it ruins the whole internet experience for me

Same here, had to configure ChatGPT to stop making these statements. Also had to configure bunch of other stuff to make it bland when answering questions.

Same, I cringe when I read this structure.

It's not text - it's clickbait distillied to grammar.

The CC point is more about the data and environmental and general configuration context, not compute and where it happens to run today. The cloud setups are clunky because of context and UIUX user in the loop considerations, not because of compute considerations.

From what I can gather, llama.cpp supports Anthropic's message format now[1], so you can use it with Claude Code[2].

[1]: https://github.com/ggml-org/llama.cpp/pull/17570

[2]: https://news.ycombinator.com/item?id=44654145

One of the most interesting coding agents to run locally is actually OpenAI Codex, since it has the ability to run against their gpt-oss models hosted by Ollama.

  codex --oss -m gpt-oss:20b

Or 120b if you can fit the larger model.

What he meant was, agents will probably not be these web abstractions that run in deployed services (langchain, crew); agents meaning the Harnesses (software wrapper) specifically that call the LLM API.

It runs on your computer because of its tooling. It can call Bash. It can literally do anything on the operating system and file system. That's what makes it different. You should think of it like a mech suit. The model is just the brain in a vat connected far away.

The section on Claude Code is very ambiguously and confusingly written, I think he meant that the agent runs on your computer (not inference) and that this is in contrast to agents running "on a website" or in the cloud:

> I think OpenAI got this wrong because I think they focused their codex / agent efforts on cloud deployments in containers orchestrated from ChatGPT instead of localhost. [...] CC got this order of precedence correct and packaged it into a beautiful, minimal, compelling CLI form factor that changed what AI looks like - it's not just a website you go to like Google, it's a little spirit/ghost that "lives" on your computer. This is a new, distinct paradigm of interaction with an AI.

However, if so, this is definitely a distinction that needs to be made far more clearly.

On the other side of the spectrum, I see some of the latest agents, like Codex, take care to get accessibility right -- something not even many humans bother to do.

But that's exactly what an LLM solved.

It's the best ui ever.

It understands a lot of languages and abstract concepts.

It will not be necessary at all to let LLM generate random uis.

I'm not a native English speaker. I sometimes just throw in a German word and it just works.

>our favored format - in images, infographics, slides, whiteboards, animations/videos, web apps, etc

If you look at how humans actually communicate I'd guess #1 is text/speech, #2 pictures

> The distinction Karpathy draws between "growing animals" and "summoning ghosts" via RLVR

I don't see these descriptions as very insightful.

The difference between general/animal intelligence and jagged/LLM intelligence is simply that humans/animals really ARE intelligent (the word was created to describe this human capability), while LLMs are just echoing narrow portions of the intelligent output of humans (those portions that are amenable to RLVR capture).

For an artificial intelligence to be intelligent in it's own right, and therefore be generally intelligent, it would need to need - like an animal - to be embodied (even if only virtually), autonomous, predicting the outcomes of it's own actions (not auto-regressively trained), learning incrementally and continually, built with innate traits like curiosity and boredom to put and keep itself in learning situations, etc.

Of course not all animals are generally intelligent - many (insects, fish, reptiles, many birds) just have narrow "hard coded" instinctual behaviors, but others like humans are generalists who evolution have therefore honed for adaptive lifetime learning and general intelligence.

I've been doing it for months, it's lovely

https://tech.lgbt/@graeme/115749759729642908

It's a stack based on finishing the job Jupyter started. Fences as functions, callable and composable.

Same shape as an MCP. No training required, just walk them through the patterns.

Literally, it's spatially organized. Turns out a woman named Mrs Curwen and I share some thoughts on pedagogy.

There does in fact exist a functor that maps 18th century piano instruction to context engineering. We play with it

> I’m also sold on his take on "vibe coding" leading to ephemeral software; the idea of spinning up a custom, one-off tokenizer or app just to debug a single issue, and then deleting it, feels like a real shift.

We should keep in mind that currently our LLM use is subsidized. When the money dries up and we have to pay the real prices I’ll be interested to see if we can still consider whipping up one time apps as basically free

Any tips to spot this? I want to avoid arguing with a X bot.

Just guessing here, but these small models may well be essentially distillations of larger ones, with this being where their power comes from. e.g. Use a large model to generate synthetic reasoning traces, then train a small model on those.

check out Sasha Luccioni

A few days ago I was trying to unsubscribe to a service (notably an AI 3D modeling tool that I was curious about).

I spent 5 minutes trying to find a way to unsubscribe and couldn't. Finally, I found it buried in the plan page as one of those low-contrast ellipses on the plan card.

Instead of unsubscribing me or taking me to a form, it opened a convos with an AI chatbot with a preconfigured "unsubscribe" prompt. I have never felt more angry with a UI that I had to waste more time talking to a robot before it would render the unsubscribe button in the chat.

Why would we bring the most hated feature of automated phone calls to apps? As a frontend engineer I am horrified by these trends.

It's probably increased during my lifetime. People used to talk, now they sit and text into smartphones.

I dont understand, the agent mode of copilot will search for and be pretty good and filling its own context afaik. I never really feed any of our 100k+ lines legacy codebase explicitly to the LLM.

Google is doing that with A2UI. LLM will be able to decide how to present info to the user.

I don't think it will become well rounded because that is not cost sensitive. Intelligence is sensitive to cost, it is the core constraint shaping it. Any action has a cost - energy, materials, time, opportunity or social. Intelligence is solving the cost equation, if we can't solve it we die. Cost is also why we specialize, in a group we can offload some intelligence to others. LLMs also have their own costs, and are shaped by it into some kind of jagged intelligence, they are no spherical cows either.

What's interesting about Nano Banana (and even more so video models like Veo 3) is that they act as a weird kind of world model when you consider that they accept images as input and return images as output.

Give it an image of a maze, it can output that same image with the maze completed (maybe).

There's a fantastic article about that for image-to-video models here: https://video-zero-shot.github.io/

> We demonstrate that Veo 3 can zero-shot solve a broad variety of tasks it wasn't explicitly trained for: segmenting objects, detecting edges, editing images, understanding physical properties, recognizing object affordances, simulating tool use, and much more.

I think he is referring to capability, not architecture, and say that NB is at the point that it is suggestive of the near-future capability of using GenAI models to create their own UI as needed.

NB (Gemini 2.5 Flash Image) isn't the first major-vendor LLM-based image gen model, after all; GPT Image 1 was first.

I agree with this fwiw, for many months I talked to people who never used o3 and didn’t know what it was because it sounded weird. Maybe it wasn’t obvious at the time but that was a good major point release to make then.

And also accessible sans login via https://xcancel.com/karpathy/status/2002118205729562949 .

I think their non-deterministic nature is what’s making it difficult to adopt. It’s hard to train somebody in the old way of “if you see this, do this” because when you call the LLM twice you most likely get different results.

It took a long time to computerize businesses and it might take some time to adopt/adapt to LLMs.

I’ve found incredible productivity gains writing (vibe coding) tools for myself that will never need to be “productionised” or even used by another person. Heck even I will probably never use the latest log retrieval tool, which exists purely for Claude code to invoke it. There is a ton of useful software yet to be written for which there _is_ no “last 20%”.

All software is not meant to be open-source, in production and working on 100 platforms.

Sometimes the point of the software is to make an app with 2 buttons for your mom to help her do her grocery shopping easier

Do you mean vibe coding as-in producing unreviewed code with LLMs and prompting at it until it appears to work, or vibe coding as a catch-all for any time someone uses AI-assistance to help them write code?

For me, that’s kind of the point. It’s similar to how the characters in a novel don’t really exist, and yet you can’t really discuss what happens in a novel without pretending that they do. It doesn’t really make sense to treat the author’s motivations and each character’s motivations as the same.

Similarly, we’re all talking to ghosts now, which aren’t real, and yet there is something there that we can talk about. There are obvious behavioral differences depending on what persona the LLM is generating text for.

I also like the hint of danger in “talking to ghosts.” It’s difficult to see how a rational adult could be in any danger from just talking, but I believe the news reports that some people who get too deep into it get “possessed.”

Consciousness is weird and nobody understands it. There is no good reason to assume that these systems have it. But there is also no good reason to rule it out.

That’s the old way of thinking about it. there is a new way.

You sound as if you have grounds for certainty about this. What are they?