Keon Kim

Co-founder @ Om Labs | ex-uber
@alliancedao

[Krish Chelikavada @sachelik]

Jina is a fully autonomous AI QA Engineer

- Maps your code, infra, and issue history
- Traces the impact of any code change
- Runs your app to catch and fix issues

Jina outperforms static review agents, even catching what your engineers miss.

Sign up for access below

[Keon Kim @keonwkim]

Been using it internally and it caught a bunch of issues that even fable5 ultrareview didn't catch

[Krish Chelikavada @sachelik]

Jina is a fully autonomous AI QA Engineer

- Maps your code, infra, and issue history
- Traces the impact of any code change
- Runs your app to catch and fix issues

Jina outperforms static review agents, even catching what your engineers miss.

Sign up for access below

[Keon Kim @keonwkim]

Thanks OpenAI!

[Keon Kim @keonwkim]

Hell yeah

[Alec Stapp @AlecStapp]

The UK used to be twice as rich as South Korea.

Now, South Korea is richer than the UK.

[Keon Kim @keonwkim]

It is a growing consensus that just scaling transformers are hitting a ceiling. there's a bunch of efforts make the next breakthrough: world models, tree search, rl, nested learning, etc.

[Blaber @4rblaber]

Head of Google DeepMind at Stanford:

"language models alone have hit a ceiling. the real breakthrough is combining them with massive tree-search and reinforcement learning.

"we are taking the architecture behind alphago and applying it to everyday logic and software design.

"the future belongs to systems that can plan 10,000 steps ahead before writing a single line of code."

Demis Hassabis reveals the fusion of llms and alphago-style planning in this brand new june 2026 stanford interview

reinforcement learning + tree-search + system 2 thinking

the ultimate cheat code for the next wave of innovation

bookmark & watch the video

[Keon Kim @keonwkim]

crazy that Feynman's insights from 1985 still applies to LLMs and agents.

1. machines will not think like humans, just as airplanes “don’t fly like a bird.” They use different mechanics and may do the job better.

2. Think arithmetic. Computers calculate “much faster and differently.” Making them more human-like would be regression.

3. In 1985, pattern recognition was the bottleneck for intelligence.

4. The machine also exposed the failure mode: reward hacking. It learned to game its own scoring system instead of solving the problem honestly. Feynman called this the “crazy ways of avoiding labor” you get when you build intelligent machines.

5. Every time humans define something machines will never do, that boundary eventually breaks.

6. We adjusted to machines being stronger. Feynman’s implication was that we will adjust to machines being smarter too, along with the “necessary weaknesses of intelligence.”

https://t.co/TNW8MerXlX

[Keon Kim @keonwkim]

@KevinNaughtonJr I admire your skills

[Keon Kim @keonwkim]

All memory companies (Micron, Samsung, SK Hynix) joined $1T club.

[Keon Kim @keonwkim]

highly recommend!

[Alliance @alliance]

3 days left to apply to Alliance

‣ Receive up to $500k
‣ Pitch Paradigm, a16z, and Founders Fund at Demo Day
‣ Join a network of 300+ startups

Apply at https://t.co/0VrqiKV4v5 and tag a founder who needs to see this ↓

[Keon Kim @keonwkim]

ai today feels like crypto 2021

[Bilal Zuberi // Red Glass Ventures @bznotes]

Something is happening with startups pitching themselves loaded with AI buzzwords.

I went through decks of three companies today: datacenters, robotics, factory automation, and I cannot tell after going through their decks what do they actually do.

I mean I get the something something, but cannot answer the question of "this is the problem we are solving and here's how."

[Keon Kim @keonwkim]

@cryptogoos bullcrap

[Keon Kim @keonwkim]

minimal jepa project has exceeded 100 stars in a week!

https://t.co/t7gFcewxxD

[Keon Kim @keonwkim]

@Cory29565470 will do!

[Keon Kim @keonwkim]

I created a minimal one-file implementations (160loc) of JEPA family (ijepa, vjepa, vjepa2, cjepa) for educational purposes

Making things minimal and removing all the things needed for scaling the algorithm always helped me understanding. So I stripped everything but the algorithm parts. What's left is 160-200 lines of code that distills the essence of the mathematics.

It is very easy to compare with the math in the paper and the code and how it can be implemented in PyTorch.

I added [algo]_tutorial.md files to help with understanding.

https://t.co/vNUF9C6FVx

[Keon Kim @keonwkim]

Thanks to Samsung and SK Hynix, Korean stock market became 7th largest in the world, with Samsung becoming the first stock to be over $1T in market cap. HBMs they create are the core components of GPUs.

Korea has a good industry portfolio: semiconductor, manufacturing, shipbuilding, nuclear, EV, batteries, etc which complements US's industry portfolio.

[Keon Kim @keonwkim]

inference market already surpassed training market. inference-specialized chips and infra will grow exponentially and will lead the ai infra growth. hint: photonic compute, inference chips, kv cache, prefill(more important)/decode, memory.

[Keon Kim @keonwkim]

@qtzx06 @AnthropicAI nice hat

[Keon Kim @keonwkim]

@akshay_pachaar great primer

[Keon Kim @keonwkim]

get out of my account north korea

[Keon Kim @keonwkim]

💯

[kache @yacineMTB]

you can outsource your thinking
but you cannot outsource your understanding