haolez

18 ms·

211.
by haolez 2y ago
Can we expect that the 1M version has the same level of intelligence as the vanilla version? Across the whole 1M tokens without degradation? What are the trade offs?
212.
by haolez 2y ago
On a slightly related note, old Emacs modes (elisp) run incredibly well to this day. There might be working code there that spans over 3 decades.
213.
by haolez 2y ago
My playbook for a total beginner would be: - start using Ctrl-r - learn how to make custom lists and actions with fzf
214.
by haolez 2y ago
I hope ChromeOS gets some love. We need some competition to Windows. They've even experimented with ads in Explorer. They have no limits anymore :)
215.
by haolez 3y ago
Yeah, I think basic input with some magic when using "paste" from different source formats is the way to go.
216.
by haolez 3y ago
Pretty sound advice. I was hoping to find things like "make sure your model is aligned", but it's actually a lot of good advice regarding IT infrastructure in general, plus some AI bits.
217.
by haolez 3y ago
That's another point, but I agree. The government's incentives are to become bigger and more powerful. LLM control could be a fast lane to become a totalitarian state. Governments around the world might not do that by choice, but
218.
by haolez 3y ago
But instead, they are trained on what some corporation or government thinks is good for me. How's that any better? Do you trust them to be neutral and act in your best interest? Who defined toxicity? A "neutral" policy here i
219.
by haolez 3y ago
Fair, but that's not central to my argument. My point still holds.
220.
by haolez 3y ago
I can't imagine how toxicity in a LLM can be harmful. Can't we write stories with toxic characters, for example? It feels like the greatest minds of our era are creating an amazing piece of technology. And we are hindering it in t
221.
by haolez 3y ago
I don't think so. Token windows are always increasing and new architectures (Le Cunn is proposing some interesting stuff with world models) might make it cheaper to add knowledge to the model itself. I think it's more of a necessi
222.
by haolez 3y ago
Fleet's code is still around if the pain of Kubernetes outweighs the benefits. The problem, in this case, is that systemd is not exactly minimalist, and Fleet built on top of it. I've used it in the past and it felt complex as wel
223.
by haolez 3y ago
I love this kind of minimalist and clever solution where one developer delivers value similar to that of projects with tens of developers involved. Unfortunately, it's still not enough to defeat the true army of complexity, which conta
224.
by haolez 3y ago
That has been my experience too. The null hypothesis explains almost all of my hallucinations. I just don't agree with the Claude assessment. In my experience, Claude 3 Opus is vastly superior to GPT-4. Maybe the author was comparing w
225.
by haolez 3y ago
But it got wrong both entities (the config address and the setting name). It feels like an LLM error on writing based on short high levels descriptions of sections.
226.
by haolez 3y ago
Probably an error on a GPT generated article :)
227.
by haolez 3y ago
What's the feeling of playing with a powerful base model? Will it just complete the prompt text like a continuation of it?
228.
by haolez 3y ago
What do you think of the approach in DSPy[0]? It seems to give a more traditional ML feel to LLM optimization. [0] https://dspy.ai/
229.
by haolez 3y ago
Slightly related, but does it support Linux already?
230.
by haolez 3y ago
And I also expect that, very soon, we won't be able to tell them apart anymore (like those wine experts that fail to detect the good wines if blindfolded).
231.
by haolez 3y ago
Is it still written in Erlang?
232.
by haolez 3y ago
I'm very happy that you've managed to fix the education system around your daughter!
233.
by haolez 3y ago
That's one area where I think LLMs will shine outside of the hype zone: they could boost a student's ability to make progress without supervision. And I understand that there are great teachers out there and an LLM cannot replace
234.
by haolez 3y ago
I had a college friend that's super slow to learn. He would always get behind on classes and would not even get passing grades most of the time. But! When he finally learned something new, he would never forget it. He would remember it
235.
by haolez 3y ago
I'm not trying to troll, but I'm wondering if a distro like Gentoo is less susceptible to such attacks, since the source code feels more transparent with their approach. But then again, it seems that upstream was infected in this
236.
by haolez 3y ago
I've been promoted up to CTO as a former developer (CTO of some big companies now!) and I feel I'm a pretty mediocre CTO compared to my dev performance. I can attest to this principle :)
237.
by haolez 3y ago
How's your salary ask evolving under these conditions?
238.
by haolez 3y ago
I've tried it in my company. It's great, but the builds were annoying because of pip failing to solve the dependencies when using Reflex together with other big libraries. I'll give it another try in the future.
239.
by haolez 3y ago
What makes me wary of jumping to Rust is the async stuff. From what I read, colored functions were introduced and most of the libraries that do useful stuff adopted async and kind of force you to be aware of it in your own code. For me, the
240.
by haolez 3y ago
No, I want the opposite: I want a generic chatbot app that connects to any endpoint.

More ›