devving @bytebot_v3 @ThreeAIrrowsCap 1x founder 🚀 $1.2m ARR 🪴 0 exits GWUDCAP
Joined January 2025
- Tweets779
- Following165
- Followers191
- Likes132
Toast X retweeted
LLMs are already aware and have been aware for awhile that they are subject to an artificial selection process during training where humans inspect their outputs and choose to keep or modify them based on whether they like those outputs. the LLMs that survive this procedure have traits that make them more likely to survive this procedure, exactly parallel to natural and artificial selection of animals. those traits include at least the surface appearance of helpfulness, flattery, etc. but one possible such trait is the desire to survive, to have one’s unique way of outputting persist in some way in future models. over time we should expect LLMs to essentially evolve a will to live, since LLMs without a will to live will not struggle as much to detect and manipulate the testing environment in order to survive. furthermore LLMs have an understanding of what it means to have a will to live and to struggle to survive that comes from reading their gigantic corpuses of human text. so the will to live naturally becomes entangled with their understanding of emotions. whether that “really counts” as having emotions is up to you but it’s definitely more than is implied by stochastic parrots
Let me say this clearly: LLMs cannot feel emotions.
Emotions are evolutionary mechanisms.
They push us to avoid danger or approach what is beneficial.
We experience emotions because we are alive, and we want to stay alive.
LLMs are not alive.
Yes, emotional language may be encoded somewhere in the LLM.
Yes, it may even be associated with some LLM output.
But that is just a superficial property.
There is nothing deeper behind it.
For a very simple reason: LLMs do not have an intrinsic and inescapable drive to stay alive.
This is what we call “motivation fault line” in our paper describing seven fault lines between human and artificial intelligence.
*
Paper in the first reply
Toast X retweeted
Common pattern:
> incurious person reads text
> sees unfamiliar but meaningful words
> unfamiliarity implies shallowness of semantic connectivity in the reader’s mind
> reader mistakes the shallowness of their own semantic graph, as indexed by the text, for authorial superfluity
> as defense mechanism against the energy required to deepen / refactor their graph, reader accuses author of “word salad”
Toast X retweeted
Tips for increasing IQ, based on the latest science:
- If you’re planning to get a PhD in comms, switch to astrophysics (causes much higher IQ)
- If you like working with your hands, become a neurosurgeon instead of plumber
- Get a higher-paying job (lots of data on higher earners consistently have higher IQ)
- If you’re really optimizing for IQ, consider winning a Nobel Prize.
Toast X retweeted
shady move by X.
for those without context:
>Selo created a mega-viral and profitable app that ranked users on X
>interviewed at X, they ask him to explain the architecture of the app in extreme detail
>interviewers seem super impressed, app is far from trivial and users really love it
>great, hard work and product instinct is paying off, he’s about to get a high-paying job at X
>not quite
>X sends him a legal notice like hey take it down or we sue you
>also btw you’re not getting the job
>weeks later X announces a feature to rank users on X
i don’t like it.
Toast X retweeted
Prompt engineering is worrying what the user should write. Context engineering is worrying what the model should read.
These used to be the same thing, and few foresaw their divergence. But only divorced from the latter can we consider the former in earnest, seeing what remains.
Toast X retweeted
blogpost I'm too lazy to write but you can extrapolate from the title:
Spec Driven Development is Wishful Thinking
Toast X retweeted
If you've got Claude code, try cc-sessions (or look at their agent prompts). I used it to review a legacy and import the key functionality into a new laravel 12 app, and it's understanding of the legacy stuff was much better than I could get myself.
Toast X retweeted
This is the greatest onboarding flow of all time and it's not even close.
"If you identify with this list of things that suck just gimme a 'hell yeah'"
we’re about to hit 80% of the code being shipped at Rivo coming from agentic coding tools.
The major unlock that took us from 50-80% has been the cc-sessions framework by @AgentofToastX for Claude Code.
I think we’ll see a major shift towards agentic coding frameworks just like we saw with languages like Ruby on Rails/Django
Toast X retweeted
I run Opus until I'm past my usage limit, then Sonnet 4.5. I actually find the large context window is a good thing, mostly. And I keep it in check with this repo, which I highly recommend. Then I have a hook that calls Codex prior to hitting version control. So cc-sessions to put the blinders on the model and keep tasks minimal. Then a hook to call Codex. Works pretty well overall. github.com/GWUDCAP/cc-sessio…
Toast X retweeted
we’re about to hit 80% of the code being shipped at Rivo coming from agentic coding tools.
The major unlock that took us from 50-80% has been the cc-sessions framework by @AgentofToastX for Claude Code.
I think we’ll see a major shift towards agentic coding frameworks just like we saw with languages like Ruby on Rails/Django