FleetingBits
@fleetingbits
Followers
1K
Following
44K
Media
1K
Statuses
7K
What's causing the slow development and adoption of audio models (like Suno) compared to image models?.
6
0
9
Some basic principles that I have about AI that have helped me to avoid hyped research:. 1) Compute is an upper bound on lab capabilities; if a lab puts out a model much better than its compute would imply, it's overhyped (e.g. Reflection 70B) . 2) The important discoveries are.
2
1
24
AI is going to allow a $700/hr lawyer to compete with a $1,400/hr lawyer. We are not there yet, but we are close, and it's going to have a huge effect on the legal market. The distribution of legal salaries is going to become a lot more bimodal over the next 5-10 years.
3
0
20
One of the most important things that can be done now to improve safety at frontier labs is to ensure that safety auditing has: (1) security of tenure, (2) pay equal to capabilities and separate from lab success, and (3) a direct line to the government and the board.
Many people share the misperception that the solution to the existential dangers of advanced AI is that individuals silently refrain from using AI. It's more that everyone needs to demand regulation of frontier AI companies from their political representatives.
2
0
11
Almost got emotional when I received that swag from @ycombinator today! . (The t-shirt we received when we started our batch in 2018 was « build something people want »). What a ride since then
0
0
4
What does it mean for an LLM to suffer? Where is suffering in the activations / outputs of an LLM?.
5
1
16
I recommend reading Albert Gu's blog post on SSMs and Transformers (linked below). The important insight for me was that transformers are biased toward sequences where you want perfect recall of all elements and SSMs toward places where you want infinite sequence length.
1
1
23
OpenAI trained their biological model using a language model base with additional training on raw biological data afterwards. This is an important insight different from previous work, which seemed to be pushing towards AI scientist or domain specific model (see Evo2).
we built GPT-4b micro (the “b” stands for bio) from a smaller version of GPT-4o - which already understands vast scientific literature - and further trained it on massive datasets of raw biological data, including protein sequences, structures and functional annotations.
0
0
11
Does anyone have good ideas for how to programmatically create random scenes (300+) of geometric figures?. Looking for diversity of scenes.
8
1
10
the problem with things like ARC having a train set is that you can probably do synthetic generation of the private set if you are willing to spend enough compute with an LLM.
1
0
5
little bit of YC lore. above a certain exit size, you get a t-shirt that says “I made something people want”.
2
0
9