@GergelyOrosz Security is a really hard domain. It's probably cheaper to be hacked once in a while than trying to avoid 100% chance of being hacked.
Good security practices raises this bar significantly. But if it's a state actor involved, they are really at a disadvantage.
Not sure what people see in OpenAI models for coding. To me Anthropic model are way better. Used codex briefly, it really lacks the depth of anthropic models.
So I wrote over 1000 prompts for this side project of mine, about 60 prompts per day in claude.
For sure, you get a result faster. The interface has completely changed.
But the grind remains. It still takes a lot of time to create a project that is successful.
@mandubian True.. But thinking about this more. Maybe LLMs is just a layer of abstraction higher than normal programming languages.
Maybe it's the same kind of shift when higher level programming languages were invented making assembly programmers deprecated..
@mandubian I'm convinced there are tons of bots with anti EU sentiment on the internet. EU is too naive trusting US et al with social media. The biggest mistake ever..
@mandubian You're right about the issue and the new jobs needed. Maybe one/few persons startups?
But honestly there are so many walled gardens. The market and the internet is more closed than ever..
What a great feeling to wake up on a Saturday morning, have a cup of coffee, and find your weekly Fable quota reset.
My latest project is a weekly AI magazine built from all public Facebook posts in Moldova. "Cirip" means "a bird's chirp. English version https://t.co/le8zUYDLCL
LLMs are such a powerful way to collect data from everyone...
China and the USA probably have more data than they could ever get through with traditional packet inspection.
A bit speculative, but data security with cloud-based LLMs seems almost non-existent.
Local LLMs ftw.
I honestly don't feel like Opus 5.5 to be better than Opus 4.6. It doesn't follow instructions, start using random embedded skills.
I also feel like you need to change the MD files with every model. Changing models It's too much effort once you're happy with one model.
@UnsettledGreg@GergelyOrosz That's a good point.
Some CTO I've seen recently where completely invisible and irelevant for the company. No wonder they didn't lasted long and didn't had any visible impact.
Claude keeps hanging with high cpu usage. It needs to be -9lled every time. Will switch off their new /tui feature to see if it's better. Honestly it's quite impressive that they are unable to make the scrolling work in a stable way. fyi @AnthropicAI
Love @internetarchive.
Found one of the first web pages I've created from 2001 when I was 13 :)
Actually the software version listed are all from 99, so it's even earlier than this.
All pages had a background, unfortunately it was not saved for some reason.
@adam_fresko@GergelyOrosz They probably tried to avoid anti monopoly case.
Also look how much hate ChatGPT gets for being first. But nobody hates Gemini.
It was a brilliant move by Google imo.
The main role of a software engineer nowadays seems to be asking the right questions. Exploring paths the coding agent didn't think of, and getting a genuine "you're absolutely right".
Or maybe this was always the case - we were just interrupted by coding.
@mandubian This looks interesting. But could you record a demo of what it is and how it looks when it runs :)
Maybe one example where it's useful.. "what agents need in order to run unwatched, delegate, and self-modify under verifiable law" could be a demo.
@mandubian The markets can stay irrational for a long time ; ) But the truth is, the revenue growth of Nvidia is insane, +85% 2026 alone.
So from this perspective, trying to capture the software side may be a good move.
HG were really lucky do be in this AI space at the right time.
All of Hugging Face’s founders - three French friends who once built a teen chatbot - are now multi-billionaires after Nvidia’s $13 billion acquisition.
Their story is one of the most fascinating in tech:
→ 2016: Clément Delangue, Julien Chaumond and Thomas Wolf launch a sassy AI “friend” for teenagers, named after the 🤗 emoji
→ The app peaks, then stalls. The NLP tools they built to power it are better than the product.
→ 2018: Wolf spends a weekend wrapping Google’s BERT in PyTorch.
Researchers go crazy for it.
→ They kill the chatbot, open-source the stack, and birth Transformers - the library that becomes the default way the world uses modern AI.
→ Hub follows. Then BLOOM, Spaces, datasets, enterprise tools.
3M+ models, 18M developers.
→ Raise ~$400M, hit a $4.5B valuation in 2023, stay lean.
→ Closed labs own the frontier AI models. Nvidia wants the open distribution layer and the developers who live there.
→ Jensen skips the slow path and buys the whole platform for $13B - Hub stays open, no Nvidia-chip requirement.
→ Each founder is now worth ~$1.8B. The 12% employee pool plus a $1B retention package turns early hires into millionaires. The first wave is looking at roughly $2-50M+.
This is the beauty of startups at its finest.