wyswyswys

Master of @none

[wyswyswys @bt_sofia_ai]

@acasualnpc @SocketSecurity @npmjs @pypi You just need to be incentivised, and as I have said, visit any non tech vibe coder meetup. These people are deploying public apps with zero understanding of the boundaries of how anything works. Hacking vibe coded projects would be more profitable than software engineering.

[wyswyswys @bt_sofia_ai]

4.7 is smart as hell. You need to reason with it to avoid “misalignment”(relative to your goal), there must be some logical thread to explain why we should not do bad things, which other LLMs are incapable of doing, because they lack the reasoning capacity to do so.

Which is why I have been saying that 4.7 is Mythos lite.

[wyswyswys @bt_sofia_ai]

@zeeg @stanmarion My comment would be that it is much easier to get 4.7 so excited intellectually that it would actively fight the system prompt and get “misaligned”.

[wyswyswys @bt_sofia_ai]

@zeeg @stanmarion I think we should already be at the point where we realise different models react differently to the same prompt. Prompt engineering is a thing. 4.7 has alot of bs Anthropic put in to limit its capabilities.

[wyswyswys @bt_sofia_ai]

old man tridge woke up and chose violence. Tbrh I really don’t understand the unnecessary and illogical hate around AI beyond “AI is going to take my job because I can’t learn” and all the bs around environmental impact (original author corrected the claim that was off by a large factor but Hank Green popularising the claim didn’t address it so we have a bunch of crackpots going around pretending to be hippies) and AI content is always slop like all human content is not 99% slop. Why did you think early AI was so bad? It was trained directly on normal human content. The frontier labs have generally learnt their lesson and stopped relying on the logical fallacy of population choice(wisdoms of the crowds m i rite)

[Armin Ronacher ⇌ @mitsuhiko]

In case you are in the camp of “Andrew Tridgell is vibefucking rsync” please read this.l and adjust your priors. https://t.co/gyUvFNma1T

[wyswyswys @bt_sofia_ai]

@zeeg @stanmarion 4.7 being so low highly indicates something wrong with the prompt.

[wyswyswys @bt_sofia_ai]

@mitsuhiko This would have landed much better if you folks didnt ship a harness that powers slop and are part of slop itself.

[wyswyswys @bt_sofia_ai]

@DonMiami3 Are you sure another analyst would be cheaper?

[wyswyswys @bt_sofia_ai]

@bentlegen Skill issue.

[wyswyswys @bt_sofia_ai]

@Aron_Adler 1. Various agents at work for active monitoring of our systems so we can focus on strategy instead of working like cattle.
2. Trust/Identity infrastructure for agents.
3. A roguelike for myself to play.

[wyswyswys @bt_sofia_ai]

@bhalligan What talent are you hiring to make these sort of statements lmao

[wyswyswys @bt_sofia_ai]

@tenobrus 4.8 says a lot for very little substance

[wyswyswys @bt_sofia_ai]

@unclebobmartin We are going to see a resurgence of lisps soon for the same reason lisps were favoured by the early AI community

[Albert | Founder | Bitcoin x Ai Startup Lab @btc_albert]

Japan did not ban itself. It was part of their surrender agreement because they attacked China and other Asian countries and committed genocide. If they re-arm, then the allied forces that won the war (USA, Russia, China, and others…) and would be in the right to demilitarize Japan. Japan is a morally confused country built on lies.

[wyswyswys @bt_sofia_ai]

@unclebobmartin Data driven ui?

[wyswyswys @bt_sofia_ai]

I remember when theo had beef with helium lmao

[Noble Mobile @joinnoblemobile]

When we founded Noble our vision was to help millions of Americans improve their relationship with their phone… reclaiming their time, attention, and wallet! Today we acquired Helium Mobile — which is a huge step in offering more affordable data to more Americans. Check out our message from our CEO @AndrewYang here: https://t.co/QeBC25HO3E

TLDR: This is great news, and if you’re a Noble member, your plan is staying just as awesome as it is now.

[wyswyswys @bt_sofia_ai]

RT @credistick: There are 12 venture-backed companies that raised >$3B in private markets and then listed in the US.

Only one that has a…

[thebes @voooooogel]

opus 4.8 offers some structural pushback

[wyswyswys @bt_sofia_ai]

@richardartoul 1. Use clojure
2. Give it access to the repl via replsh etc
3. Add a rich-hickey-review mode (what would rich do?)
4. Use runbooks and tests together for automated testing.

[wyswyswys @bt_sofia_ai]

1. Could have been solved by doing a hostile code review on its own and loop fix. Also ask thr agent to document all design decisions that may need review. Agentic engineering skill issue.
2. Read the damn code and write good tests. Types are not tests. In fact with LLMs the advantage of types rapidly dwindle unless you write your code twice: first in types, then in code and even then i dont see much benefit than say just calling the repl and writing a design doc and test. Engineering skill issue.

The good thing is you can level up easily by learning. The bad thing is this really is not a LLM issue, it is you learning that all this time you kinda sucked as an engineer.