Hacker Newsnew | past | comments | ask | show | jobs | submit | Tenoke's commentslogin

Claude Code has been wonderful for work and the frequent improvements are nice, although with Mythos being used by others ages ago and new versions for the public still being bellow that, it's hard to not feel like the underclass already.

It's also at minimum baked into the system prompt of virtually any LLM.


That's not "baked" and only applies to remotely hosted LLMs where someone else feeds the prompt into the LLM.


Compared to gemini-cli which I was using the last few weeks it also doesn't:

1. Doesn't tell you your weekly qutoa (at least on Pro plan/all the time)

2. Your agent cant access the quota to not run some tasks at low quota

3. You cant see the context size

4. Your agent can't see the context size

5. You can't compact/compress

6. You have to keep starting new chats which also kill any processes it has running (e.g. a telegram listener)

7. Doesn't have a straightforward linux/wsl install (I ended up using the Windows IDE and pointing it to wsl).

And that's from just migrating a gemini-cli model and trying to set it up for an hour. Incredible downgrade for no reason.


No compacting?????????


Nope, it does some automatically, but you cant even check the context size let alone compact. The agent proposes to start a new chat when you think it might be high, and that's it.


It does auto compact


It depends heavily on what type of data though. As far as I understand if you have no PII or anything close to it you are mostly safe - especially if it's customer data but aggregated.


> We did get a certificate though.

As someone who never bothered to get any certificates (beyond a University degree) even when I'd do online courses (of which the most course-like must've been fast.ai), are these ever actually useful in any manner?


There are so many stupid "courses" and SaaS tools doing this I imagine that the value of the many of these certificates is close to 0.

Many of them you can simply take the exam over and over until you pass, and then stick a shiny stupid badge on your LinkedIn profile.


They are useful for getting a job, that’s about it.

In our case, we get our entire team AWS solution architect certs as well just so we can always tell our customers that our whole team is certified (we do a lot of “forward deployed” stuff for enterprise customers).


In his case of a large company, I’d expect that completing the useless training is necessary to get access to the tool. That’s how it worked in mine.


As someone working in a small business/startup, who finally got the team Claude Team Premium, I don't really get what might I benefit extra from by enabling this. I can find whatever workflows and tell it to integrate them anyway, why would I bother with this?


They dont know :)


In 2019 I suggested[0] you might reach AGI if you train on computer usage - mouse movement, keypresses, what's on the screen etc. - and it sounds like Meta are kind of trying some form of it.

0. https://svilentodorov.xyz/blog/human-imitating-task/


A lot of people had this idea. You’re going to have to take a vague idea to fruition if you want the props you’re looking for.


Can you point me to them? I couldn't find anyone writing a version of that idea back then, I'd be curious to read how others framed it.


Its style of writing text is very readble if aesthetically meh. This is what I care for in how code is written anyway.


A great example of how current alignment is imperfect and bound to miss random behaviors nobody is trying to get.

This is cute now, and a huge problem when future AI does everything and is responsible for problems it isn't even directly optimized for. Who knows what quirks would arise then.


I think eventually you are going to end up with every smart AI continually checked by dumber AI's to make sure they don't do anything too crazy. Which probably does bring AI closer to how human intelligence works


New technology isn't perfect now -> drop technology and never use it in the future


What are you even responding to?


Completely agree, top down “alignment” and RLHF is actually quite primitive and uses a lot fancy words to describe what is essentially just hitting the machine with a stick without the nuance, context, or feedback to help it model why the feedback was given.

Also to be honest I think OpenAI models struggle a lot with this, I primarily stopped using them in the sycophancy/emoji era but ever since the way they talk or passive aggressively offer to do something with buzzwords just pisses me off so much. Like I’m constantly being negged by a robot because some SFT optimized for that really strongly to the point it can’t even hold a coherent conversation and this is called “AI safety” when it’s just haphazard data labeling


I think LeCun has been so consistently wrong and boneheaded for basically all of the AI boom, that this is much, much more likely to be bad than good for Europe. Probably one of the worst people to give that much money to that can even raise it in the field.


LeCun was stubbornly 'wrong and boneheaded' in the 80s, but turned out to be right. His contention now is that LLMs don't truly understand the physical world - I don't think we know enough yet to say whether he is wrong.


Could you please elaborate on what he was wrong about?


He said that LLMs wouldn't have common sense about how the real world physically works, because it's so obvious to humans that we don't bother putting it into text. This seems pretty foolish honestly given the scale of internet data, and even at the time LLMs could handle the example he said they couldn't

I believe he didn't think that reasoning/CoT would work well or scale like it has


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: