Jonathan Liles

116 posts

Jonathan Liles

@thjonml

Founder. Author of many open source projects. Making things run orders of magnitude faster in debug mode than others do in release mode for 20+ years.

Katılım Haziran 2011

94 Takip Edilen8 Takipçiler

Jonathan Liles@thjonml·13h

MCP is very, very bad. Seems like no thought was put into it. But of course everything has evolved so rapidly and nobody seems to have time to stop and think about what a good solution would look like. I think we should all just drop it and focus on the CLI route until someone can put their head down and come up with a workable solution.

English

2.9K

dax@thdxr·13h

mcp is so tied to the idea of 1 process = 1 session there's some things in the spec that force you to spawn a new MCP server for every active session you have ridiculous overhead - this is why i kept saying it's too early to standardize

English

894

105.8K

Jonathan Liles@thjonml·18h

@DavidOndrej1 You're assuming it won't be heavily nerfed.

English

681

David Ondrej@DavidOndrej1·18h

if Fable drops tomorrow I'm gonna lose my shit

English

294

13.4K

Jonathan Liles@thjonml·19h

There's more going on with the subscription vs API issue. For one thing, it doesn't seem viable to use an LLM at an API rate without the provider offering some guarantee that it won't go into a loop and rack up hundreds of dollars in charges in a few minutes. Once you've paid $$$ for a screen full of newlines or "Let me do... I'm about to... Going to do it now..." you never want to leave the security of a subscription again.

English

252

dax@thdxr·19h

there's something weird about everyone saying how productive their token usage is but being entirely dependent on subscriptions and completely incapable of paying per token is it providing returns or not?

English

409

34.5K

Jonathan Liles@thjonml·21h

@ThePrimeagen @neetcode1 The fact that LLMs can generate infinite permutations on meaningless marketing fluff type language isn't helping the situation. Some of it is quite egregious, possibly criminal, in terms of obfuscation.

English

164

ThePrimeagen@ThePrimeagen·22h

@neetcode1 no one has any idea what the hell they are talking about because we really don't have a solid idea of what the hell we are doing :)

English

137

4.5K

NeetCode@neetcode1·1d

> "takes a while to wrap your head around" Really? It's progress, but I feel like this isn't that new of an idea. We've been talking about slack integrations, llm context and collaborative experiences forever. Also, this looks like an example of prompting. I thought prompting was dead. Idk, I feel like everyone's just making it up as they go.

Andrej Karpathy@karpathy

This is a new paradigm for interacting with Claude that is significantly more "inline" with all the other human activity org-wide. Once you do all of the under the hood engineering work to make this "just work" (e.g. across tools, integrations, compute environments, memory, security, etc.), Claude basically joins the team in a seamless way - you can talk to it as you would talk to a person and it can help with a very large variety of workloads. Imo this is the 3rd major redesign of LLM UIUX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download to your computer. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans. It really takes a while to wrap your head around it, but it works and it is awesome.

English

632

51.2K

Jonathan Liles@thjonml·21h

This is a real annoyance/problem with using agents and git together. 1) they tend to create a lot of useless docs if not prompted out of the habit and 2) they also tend to commit them. And it's not that one wouldn't benefit from having a "PRD" tracked in git... But it shouldn't be in the public repo. This comes up with other kinds of files too (.env being a well known example). You don't want to lose it, but you don't want to commit it either, and there's no straight-forward/easy way to deal with this in git.

Matt Pocock@mattpocockuk

There are a class of AI Coding assets that IMO don't really belong checked into git: - PRD's - Research files - Decision maps - Implementation plans Folks who agree with me, what are you using instead?

English

Jonathan Liles@thjonml·21h

If you're not having the LLM write the commit message, you're not doing it right...

ThePrimeagen@ThePrimeagen

when you are vibe coding... is there really any commit message that is useful?

English

Jonathan Liles@thjonml·1d

@jun_song @grok explain why this is false.

English

463

Jun Song@jun_song·1d

Chinese AI models have one massive advantage: Their training data is mostly in Chinese, and a single Chinese character packs significantly more meaning than an English letter. This means they can compress a lot more data per token. It gives them up to a 4x token compression efficiency—an inherent advantage that English-based US models simply cannot replicate. This is exactly why GLM, a mere 750B model, can compete with 2T-level frontier models.

English

243

1.1K

125K

Jonathan Liles@thjonml·1d

@DavidOndrej1 You'll spend all night waking up suddenly and all day on the toilet...

English

130

David Ondrej@DavidOndrej1·1d

my goal is to overdose on creatine

English

5.9K

Jonathan Liles@thjonml·1d

@badlogicgames It's that 20 minutes in a block or 40 30 second quick checks while waiting for a generation?

English

184

Mario Zechner@badlogicgames·1d

setting myself a 20 minute twitter cap per day. very healthy. much recommended.

English

287

11.9K

Jonathan Liles@thjonml·1d

I think I know part of what's going on here, and it's only partly a matter of popularity...

Howaboua@Howaboua

So the Openrouter rankings... Where do you think this Hermes domination is coming from? I am not asking about whether it' a good app or not - I am wondering how the heck do so many people use it via Openrouter? Or is it just because Hermes/Kilo/OC are super token hungry? Where's Opencode?

English

187

Jonathan Liles@thjonml·2d

Most code in the training set doesn't enforce invariants. Also the LLM has no understanding of your codebase. Just imagine you were a junior coming in and asked for change something in one file in a codebase you have no understanding of, would you start adding asserts knowing that it could blow up in your face?

English

947

ThePrimeagen@ThePrimeagen·2d

Something I haven't figured out, though I haven't put enough time into it, is how to get an llm to produce good asserts for a program. It really just doesn't seem to understand this.

English

366

45K

Jonathan Liles@thjonml·2d

@ThePrimeagen Prime, if you are busy hyping your own agent, how are you going to hype my agent?

English

2.4K

ThePrimeagen@ThePrimeagen·2d

day 1 of new agent harness i have thrown together over the weekend... HEAVILY SLOPPED IT btw Anyway, I wanted to just try it out... and my gosh... I really love this

English

786

109.2K

Jonathan Liles@thjonml·2d

@opencode Not really sure what I'm looking at but it is pretty.

English

470

OpenCode@opencode·2d

we've added unique user rankings some models are token heavy so they skew upwards in rankings - unique people using the model is a more accurate ranking we'll orient more of our data around this metric

English

1.8K

117.4K

Jonathan Liles@thjonml·2d

Also very revealing about certain peoples' biases and intelligence levels. MRI is a lot more invasive, expensive, and dangerous than people apparently realize. Midjourney had its moment and (apparently) rather than spending their excess profits on a dead-end or a superyacht, they put it into something potentially very useful and beneficial. That is highly unusual.

English

3.5K

ThePrimeagen@ThePrimeagen·2d

the midjourney pivot is so much crazier than you realize

English

119.9K

Jonathan Liles@thjonml·2d

@thdxr There's a a huge power imbalance between an established company with capital to throw at something and a startup with no resources.

English

391

dax@thdxr·2d

most startups have good ideas but fail to figure out how to package them into a successful product so they mostly exist to feed ideas to the companies who have figured this out

English

619

37.1K

Jonathan Liles@thjonml·2d

I find the price of RAM skyrocketing to be a little bit of a hopeful signal for software quality actually. People have been totally reckless with this metric for decades now with basically no penalty.

ThePrimeagen@ThePrimeagen

software devs in the 70s> "How many bytes are we using? we cannot possibly afford one more byte" software devs today> "no mistakes" wild change over the last 60 years

English

Jonathan Liles@thjonml·2d

I think there may be an important lesson here about shipping early. A lot of people, myself included, tried OpenClaw but just couldn't get it to work or it maxxed their CPU, ate all their RAM, etc. I'm sure it's improved in those respects, but people moved on. When there are a 1000 clones of your project claiming to be less bloated, that's a signal (but also a lot of clout chasing) Hard to disentangle the hard effects from all the hype/social effects, but still worth thinking about. What difference in staying power would a higher quality product at launch have had? We've been having the bloated vs light argument in software for a long time (at least I have), but this is a new era and the potentials are magnified by the sheer amount of code LLMs can crank out effortlessly.

Peter Steinberger 🦞@steipete

People here discussing what happened with OpenClaw. The hype died down. We improved quality and grew a team. We created a non-profit whereas competiors are VC funded and have other agendas. This is our strongest week so far.

English

Jonathan Liles@thjonml·2d

You could theoretically route between well characterized models at the provider side... keep the TTL on the cache longer, but this is clearly just hype with no practical use. There's a bit of a "business practice" going on in the industry with encouraging people to spend a lot of tokens/money "benchmarking" the new thing (boiling the oceans to make the same threejs games over and over). Provider makes money regardless of the usefulness on the "benchmarking" alone...

dax@thdxr

i have low conviction on model routers - very open to changing my mind but this is a snapshot of my current thoughts - i don't think it's good to not be aware of what model you're using. coding with LLMs is a skill you develop and getting a feel for models is part of that - people (at scale) don't have this skill right now which is why a lot of companies are complaining that people are using expensive models for dumb things. a model router promises to solve this without the user having to do anything but i think the issue is missing feedback loops to the user. id rather we figure out how to help users get smarter - i dont even know how much you can model route when factoring in things like prompt cache. only so much you can do - their effectiveness is a bit exaggerated by the same dynamic that's impacting everything AI. so many companies desperately searching for opportunities and trying anything. model routing is the one thing models labs cannot do so everyone is jumping on it

English

Jonathan Liles@thjonml·2d

@mitchellh I recall having a fellow developer once tell me that C# was a garbage collected language and therefore memory leaks were impossible. He actually believed it, and I don't think he found my explanation convincing. I just continued fixing the leaks... Valgrind, although extremely useful, can give people the same false sense of security in C, C++, Zig, etc. One nice thing about Zig its that it's more explicit about the fact that keeping track of allocations is your concern, not the compiler's. Hint: when there's a image like this: it's a meme, not a bug report.

Kai Fall@kai_fell

written in zig btw

English

Jonathan Liles@thjonml·3d

@MiaAI_lab @NVIDIAAI Like what for example?

English

252

Mia@MiaAI_lab·3d

The amount of actual useful work I'm able to do on my @NVIDIAAI DGX Sparks is incredible and improving every day. These little machines are capable of serious work.

English

4.8K

Keşfet

@DavidOndrej1 @ThePrimeagen @neetcode1 @jun_song @grok @badlogicgames @elonmusk @BarackObama