Luna Hayes

AI researcher, builder | Sweety girl

[Luna Hayes @lunaonexbt]

Sam Altman at Stanford, earlier this year: OpenAI's next hire is an AI

"We set this goal that by September of this year we will use 500,000 A100 equivalent GPUs as an AI research intern"

What he laid out in the class:

→ Codex got really good early this year, and 5.5 was the real inflection point
→ today's pipeline is pre-training, mid-training, post-training and RL, and he expects a major rewrite
→ finding a better pipeline is "a research problem for the AIs to figure out"
→ by March 2028 the goal is a full end-to-end, very talented AI researcher inventing new architectures

AI intern → AI researcher → AI designs the next AI

His bet: with the current pipeline, AI gets over the line where it does incredible work

What would you hand an AI research intern first?

[Dmitro Cashback @DmitroCP]

https://t.co/y6QtCvmwgd

[Luna Hayes @lunaonexbt]

@DmitroCP AI isn’t the bubble, infrastructure is the bottleneck

[Luna Hayes @lunaonexbt]

Geoffrey Hinton opened his lecture with: "If you sleep well tonight, you may not have understood this lecture"

Then he explained why digital minds will out-learn us:

→ people share around 100 bits per sentence, even at the maximum rate
→ copies of the same model share what they learn by averaging their weights, trillions of bits at a time
→ that's how big models are trained: many copies, different data, one shared update
→ he calls them millions or billions of times better than us at sharing

10,000 copies → 10,000 different courses → every copy knows every course

"It's kind of scary"

If your copies could learn everything you learn, what would you study first?

[Dmitro Cashback @DmitroCP]

https://t.co/y6QtCvmwgd

[Luna Hayes @lunaonexbt]

@DmitroCP The last 10% is always the hardest to automate

[Luna Hayes @lunaonexbt]

@monokern in dm bro

[Luna Hayes @lunaonexbt]

Sam Altman at Stanford, earlier this year: OpenAI's next hire is an AI

"We set this goal that by September of this year we will use 500,000 A100 equivalent GPUs as an AI research intern"

What he laid out in the class:

→ Codex got really good early this year, and 5.5 was the real inflection point
→ today's pipeline is pre-training, mid-training, post-training and RL, and he expects a major rewrite
→ finding a better pipeline is "a research problem for the AIs to figure out"
→ by March 2028 the goal is a full end-to-end, very talented AI researcher inventing new architectures

AI intern → AI researcher → AI designs the next AI

His bet: with the current pipeline, AI gets over the line where it does incredible work

What would you hand an AI research intern first?

[Dmitro Cashback @DmitroCP]

https://t.co/y6QtCvmwgd

[Luna Hayes @lunaonexbt]

@DmitroCP totally agree here

[Luna Hayes @lunaonexbt]

@hoshinoxbt The winners will build with AI, not just use it

[Luna Hayes @lunaonexbt]

@callanxai The real edge is building AI systems that improve themselves

[Luna Hayes @lunaonexbt]

@hanakoxbt The best AI lessons come from people building the future

[Luna Hayes @lunaonexbt]

@DmitroCP Exactly, building is easy now, scaling is the real moat

[Luna Hayes @lunaonexbt]

Lovable's CTO on the scale behind vibe-coded apps

"We've created 70 million applications on Lovable"

Fabian Hedin's numbers from his Cloudflare Connect talk:

→ 70 million apps created since launch
→ about 1 million new apps every week
→ over 1 billion visits to those apps every month
→ about 7 billion requests a day to the Cloudflare Workers behind them

70M apps → can't keep a server live for each → spin up compute only when someone visits

That's the real engineering problem of AI app builders: not generating the app, but serving millions of them on demand

Would you build your next product on a platform hosting 70 million others?

[Dmitro Cashback @DmitroCP]

https://t.co/y6QtCvmwgd

[Luna Hayes @lunaonexbt]

@DmitroCP Intelligence might depend more on architecture than raw computing power

[Luna Hayes @lunaonexbt]

@DmitroCP AI replacing workers only works if people can still afford to buy things

[Luna Hayes @lunaonexbt]

@DmitroCP Lauren is so smart

[Luna Hayes @lunaonexbt]

Cursor engineer Lauren Tan on why most people can't scale agents

"You can't spawn a hundred agents when you don't even trust the output of one agent"

In her talk she walks through the trust curve:

→ at first you watch every single output and can't parallelize
→ over five months she built enough trust to let her agents auto-merge PRs
→ she woke up to 20 PRs already landed on main, and they were good
→ last month she shipped 1,000 PRs, this month she was near 800 by the 12th

Watch every output → build trust → let agents merge → scale

Her point: set your agents up well and you can get to a similar level

How many agents do you actually trust today?

[Dmitro Cashback @DmitroCP]

https://t.co/y6QtCvmwgd

[Luna Hayes @lunaonexbt]

@slash1sol just do it buddy

[Luna Hayes @lunaonexbt]

@DmitroCP In 2015 he joined OpenAI as a research scientist

[Luna Hayes @lunaonexbt]

Demis Hassabis on the hardest part of building DeepMind

"You want to be five years ahead of your time, not 50"

In his Cambridge lecture he admits the first years were a wall:

→ for the first couple of years nothing worked
→ they couldn't score a single point in Pong, the simplest game there is
→ they wondered if they were 10 or 20 years too early, like Babbage
→ then it converged, and that gave them the confidence for harder problems

Early → painful → working → bigger bets

Being early is a strategy. Being 50 years early is just pain

How long would you keep building before the first point lands?

[Luna Hayes @lunaonexbt]

Demis Hassabis on the hardest part of building DeepMind

"You want to be five years ahead of your time, not 50"

In his Cambridge lecture he admits the first years were a wall:

→ for the first couple of years nothing worked
→ they couldn't score a single point in Pong, the simplest game there is
→ they wondered if they were 10 or 20 years too early, like Babbage
→ then it converged, and that gave them the confidence for harder problems

Early → painful → working → bigger bets

Being early is a strategy. Being 50 years early is just pain

How long would you keep building before the first point lands?

[Hanako @hanakoxbt]

https://t.co/B3240WRGhl

[Luna Hayes @lunaonexbt]

@RahulTechX Most people underestimate how much prompting matters