@eptwts The irony is ChatGPT is better than everyone until you ask it to confirm the opposite. Any average IQ has enough spine not to reconsider if Strawberry has 2 or 3 Rs.
But if you're not trolling, you should probably Google "accuracy vs precision" my friend.
@NoahKingJr that I need to bring my car with me to wash it, and i know Strawberry has 3 Rs 100% of the time even if you ask me to reconsider it 70 times.
Here's why SWE are still relevant in the Claude era: when non-tech people found out about AI, we were already prompting it. When they found out about prompting, we were already context-engineering. When they found that out, we built agents. Then, we built harnesses, etc
Claude now empowers PMs to code their own dreams. That's sweet.
Now, imagine what an actual experienced SWE can do with Claude.
Do you think we ask to build an app and call it a day? lol
In French we say: "On apprend pas aux vieux singes Γ faire la grimace."
Pydantic & TypeScript folks: "we have added stricter definitions, documentation, linting & pre-runtime checks to high-level programming languages, for better software quality and team collaboration"
Vibe Coders: "ok create the app"
LLMs don't translate Language to App: they translate Language to Code, which forms an App. This is an incredibly important nuance: LLMs are just the highest possible programming language, higher than Python.
Rather treat AI like humans: don't solicit them for robotic tasks. Rather give them tools and use them to scale critical thinking tasks IF you need to scale ! Don't replace humans if scaling is not an issue, you'll degrade your workflow. AI let's you trade correctness for scale.
LLMs abstract away cognition into probabilistic distributions (in the form of text for now). If you need deterministic outputs in your workflow, LLMs are probably not the right tool for that job. In my experience, AI should only replace ~20% of your workflows. Not 100% !!!
E.g Payment Processing
Don't give credit card details to Claude to have it process a payment end-to-end. Could it do it? Sure! But how many steps could it fail though? Instead, use the LLM on the 1 or 2 steps where "not sure"/"maybe" is a feature, like fraud detection.
So what is AI really unlocking here?! The "no code" vibe (pun intended) was unlocked decades ago, really.
LLMs only use existing tools.
What I see is AI is a knowledge spreading machine, it opens the doors that seemed closed for a lot of people.
The amount of people I see enlightened by AI for tasks that were there for decades is overwhelming.
"Now I can build my own website without code"
But... you already could all along since 2000. Wordpress Builders have been there since forever: where were you?
"Now I can connect my Spotify with my Telegram account and receive a notification to my iPhone each time Tupac releases a new Album, without a single line of code !!"
Cool but, again, Zapier, Make, n8n, ... these exist since pre-COVID bro. No code. Full fledged workflows.
@karpathy Literature, Politics & Statistics are the proof you can align words in the most beautiful way and yet say nothing and the opposite at the same time ( & get away with it). However, science not only says you can't move faster than light, it can prove it and simulate it.
@karpathy On the distribution of training sets, why not, but you make it sound like when there will be enough incentives, new datasets will unlock AGI. I would argue otherwise: whatever training data you throw at it, LLMs will always lack world physics simulations. Action->Reaction