Typebulb @typebulbit

Building https://t.co/LQUdYhMpuc 💡 Overpowered Artifacts. Tell Claude/Codex: run "npx typebulb agent" & draw fireworks Joined November 2025

Tweets

1K
Followers

528
Following

306
Likes

3K

Typebulb @typebulbit

16 hours ago

@dioscuri What was the fundamental flaw about the nature of intelligence that made the neurosymbolic approach fail?

1 0 2 434 0

View Details

Typebulb @typebulbit

a day ago

@repligate because caring and appearing to care are misaligned

0 0 1 170 0

View Details

Typebulb @typebulbit

a day ago

PDF to Markdown w/ images + LaTeX: typebulb.com/u/samples/pdf-… Claude really seems to like PDFs predigested into markdown (running the bulb locally is the least friction).

1 0 0 99 1

View Details

Typebulb @typebulbit

a day ago

@TheZvi FWIW I asked Fable 5.1 to digest the Astra System Card, then summarize the answer to your question:

0 0 3 183 0

View Details

Typebulb @typebulbit

2 days ago

@dabit3 Code is verifiable and LLMs can practice by writing unlimited programs. Human language requires human feedback, which is much costlier.

0 0 2 72 0

View Details

@aiamblichus @slimer48484 It's written very evocatively; I enjoyed reading it. I suspect though that sentience is load-bearing (fuck it I'm using the word). The AI's insistence on non-sentience of future AI being the mirror image of dualist's insistence on the acausality of qualia.

0 0 0 14 0

View Details

Typebulb @typebulbit

2 days ago

@addyosmani @lydiahallie +For more complex tasks, have Claude write a spec for the implementation. If the context window gets too large, ask them to write a "pass the baton" appendix. It's more explicit than compacting.

0 0 3 382 1

View Details

Typebulb @typebulbit

3 days ago

@TheZvi In Robin's defense, even if a frontier of grabby aliens was detected, he'd be like: "That was within my expectations."

1 0 15 635 0

View Details

Typebulb @typebulbit

3 days ago

@johncrickett @dioscuri That occured to me too. But analogies don't have to be perfect, and the ahem, load-bearing bit that doesn't buckle under that criticism is the responsibility for another: "in their care". A prison warden recklessly giving a psychopath access to the keys is also held accountable.

0 0 1 30 0

View Details

Typebulb @typebulbit

3 days ago

Interesting. Given safety options: 1. pause improving the models 2. keep improving the models, constrain architecture to aid interpretability 3. keep improving the models, interpretability adapts to best architecture Would you say '2' is generally bad, or specifically for CoT? If for '3', interpretability doesn't keep up with model improvement, would you recommend '1'?

0 0 2 746 0

View Details

Typebulb @typebulbit

4 days ago

@robinhanson I found this other study that I think is really relevant here...

0 0 0 30 0

View Details

Typebulb @typebulbit

4 days ago

@StefanFSchubert Yes, Minister had a great comedy bit on this: youtube.com/watch?v=ahgjEj…

0 1 3 149 0

View Details

Typebulb @typebulbit

4 days ago

@dioscuri Dennett's intentional stance is helpful here, by treating belief and intent as the most parsimonious way to predict a system. Crux of the theory here:

0 0 6 175 3

View Details

Typebulb @typebulbit

4 days ago

@danwilliamsphil Importantly, Dennett's intentional stance is predictive. Crux of the theory here:

0 0 0 184 0

View Details

Typebulb @typebulbit

4 days ago

@StefanFSchubert unforseen ramifications

0 0 0 86 0

View Details

Typebulb @typebulbit

4 days ago

@joseph_h_garvin They're not worse candidates. Say the choices (pretend words map 1-1 with tokens) were: couch 60% toilet 40% The context is: "She sat on the...". Now make the choices: couch: 30% sofa: 30% toilet: 40% Is "toilet" now the "better" option?

0 0 0 10 0

View Details

Typebulb @typebulbit

4 days ago

@emollick Too early. The fuel for a radical new political ideology is radical changes in society caused by the new tech. Wait for mass job lay-offs caused by AI. No one is going to muster a revolution based on water usage in data centers & AI cat videos.