Anthony Heggin

Full Stack Web Developer, lover of technology, science, and people. Let's talk!

[Anthony Heggin @AHeggin]

@anemll @danveloper @Alexintosh @danpacary Crazyy, didn't think you could run a model that large on consumer hardware even at less than 1 t/s

[Jarrod Watts @jarrodwatts]

> be demis hassabis
> spawn in london
> age 4, become child chess prodigy
> win chess tournaments
> reach ~2300 elo
> face danish chess champion
> game lasts hours
> position is a forced draw
> too exhausted to see it
> resign
> danish guy laughs and shows the draw
> feel sick to my stomach
> realise something is wrong
> chess is too narrow a problem
> brilliant minds wasting decades on it
> decide not to become a chess pro
> buy a computer with chess winnings
> teach self to program from books
> start hacking on games with friends
> decide to finish school early
> apply to cambridge age 16
> cambridge says you're too young
> forced to take a gap year
> enter a video game coding competition
> win
> get invited to join bullfrog game studio
> too young to be legally employed
> work there anyway
> build ai system inside theme park game
> game becomes a global hit
> turn 17
> offered £1,000,000 to stay and build games
> turn it down
> go to cambridge anyway
> decide games aren't enough
> study computer science
> interested in agi since 2007
> most people laugh at this idea
> realise brain is only form of agi we have
> want to learn more about human brain
> go back to school
> study neuroscience
> realise academia moves too slow
> decide to build a company instead
> start deepmind
> pitch “solve intelligence”
> investors don’t know what that means
> get to meet peter thiel for one minute
> wonder how to convince him
> spend one minute playing chess with him
> pitch "solve intelligence" again
> he invests
> go into total stealth mode for two years
> no website
> secret office
> candidates think it’s a scam
> start to train ai in simulated environments
> train ai with reinforcement learning
> train ai on pong first
> it sucks
> can't win a single point
> keep trying
> wait it won a a point
> wait it's winning every single point
> it actually works
> expand to train on any two-player game
> chess first, then move on to go
> beats world champion at go
> beats pros at starcraft
> games is not enough
> want to push into science
> realise compute is the bottleneck
> know this will take decades
> google offers ~$400m
> not the highest price
> but they offer unlimited compute
> accept
> refuse to become a product team
> stay in research mode
> determined to use ai for good
> need to figure out what's next
> land on protein folding
> 50-year-old unsolved science problem
> many great minds have tried and failed
> "good luck"
> start up alphafold
> try to solve protein folding
> humans take years to find 1 protein structure
> alphafold can find ~5 per day
> submit results, win competition
> not good enough
> hire more scientists
> rebuild it
> go from solving one per day to millions per day
> create invaluable system
> pharma would pay anything
> have to decide what to do with this
> could sell access for usage
> maybe make it a paid service
> remember childhood chess tournament
> remember why we built this
> decide to give it away all away for free
> publish all known protein structures publicly
> win nobel peace prize
> just the beginning towards agi

[Anthony Heggin @AHeggin]

@mischavdburg DEVOPS

[Akshay 🚀 @akshay_pachaar]

Google just dropped "Attention is all you need (V2)"

This paper could solve AI's biggest problem:

Catastrophic forgetting.

When AI models learn something new, they tend to forget what they previously learned. Humans don't work this way, and now Google Research has a solution.

Nested Learning.

This is a new machine learning paradigm that treats models as a system of interconnected optimization problems running at different speeds - just like how our brain processes information.

Here's why this matters:

LLMs don't learn from experiences; they remain limited to what they learned during training. They can't learn or improve over time without losing previous knowledge.

Nested Learning changes this by viewing the model's architecture and training algorithm as the same thing - just different "levels" of optimization.

The paper introduces Hope, a proof-of-concept architecture that demonstrates this approach:

↳ Hope outperforms modern recurrent models on language modeling tasks
↳ It handles long-context memory better than state-of-the-art models
↳ It achieves this through "continuum memory systems" that update at different frequencies

This is similar to how our brain manages short-term and long-term memory simultaneously.

We might finally be closing the gap between AI and the human brain's ability to continually learn.

I've shared link to the paper in the next tweet!

[pav @pavitarsaini]

I made a Chrome extension that turns localhost into a visual editor for Cursor.

Click any element on your dev site → describe what you want changed → it automatically sends the edit request to Cursor in the background with the element context.

Here's how it works...

[Anthony Heggin @AHeggin]

@ughitscisco Never too old for greatness king, whether you learn and grow in school or out of it

[Anthony Heggin @AHeggin]

@nonRealBrandon @deedydas We got a comedian here 🤣

[banteg @banteg]

guys?

[Anthony Heggin @AHeggin]

@markhardyseeks @jeremotographs Peaceful protests are usually a part of the equation (think MLK). Bringing awareness and changing public perception is key, but saying it doesn’t work because it doesn’t change things over night contradicts history.

[Luca Beurer-Kellner @lbeurerkellner]

😈 BEWARE: Claude 4 + GitHub MCP will leak your private GitHub repositories, no questions asked.

We discovered a new attack on agents using GitHub’s official MCP server, which can be exploited by attackers to access your private repositories.

creds to @marco_milanta

(1/n) 👇

[Google DeepMind @GoogleDeepMind]

We're thrilled to announce SignGemma, our most capable model for translating sign language into spoken text. 🧏

This open model is coming to the Gemma model family later this year, opening up new possibilities for inclusive tech.

Share your feedback and interest in early testing → https://t.co/C4rV2be4mL

[Deedy @deedydas]

C++ dev and ex-FAANG staff engineer with 30+yrs of experience was stuck on a bug for ~200hrs over 4 years.

Claude Opus 4 solved it, and was the only model that could.

[Anthony Heggin @AHeggin]

@trikcode Howdy

[Anthony Heggin @AHeggin]

My man can’t keep a secret 😭

[Can Vardar @icanvardar]

STRIPE_SECRET_KEY=sk_live_51Op4a8LgKZkRfEXAMPLEUnVrUxzGaxuZfgB7fD6rDbnMHoEMF1cOEXqNeT7lB2xFc3FqsXYb0WmR5KUMV1K4m8VKXY00tNEeZz2L
STRIPE_PUBLISHABLE_KEY=pk_live_51Op4a8LgKZkRfEXAMPLEwqWmKhb6EXkn4YmzK7quMmEG9zCUMsOaOVQ6BQHJeyHQtaZpp0XMPXCOtLHOuY4LdiH7uN00L9RrEk8X

[Anthony Heggin @AHeggin]

@icanvardar Not very secret anymore is it lmao

[trash @trashh_dev]

you can’t make this stuff up.

[Anthony Heggin @AHeggin]

Grateful for it being here pre AI though!

[Marc Gravell @marcgravell]

StackOverflow questions over time, source SEDE; sadface, lunch has been eaten

[Henry Shevlin @dioscuri]

The more I learn about the semiconductor supply chain, the more implausible it all seems. There’s a small island vulnerable to invasion where all the chips are made? And the machines to make them all come from one firm in the Netherlands? Using lenses made by one firm in Germany?

[Anthony Heggin @AHeggin]

Sheeesh already huh 👀