davinci
@leothecurious
Followers
3K
Following
100K
Media
214
Statuses
8K
teaching robots to see by day, learning from nature by night. in search of elegant solutions to the metaproblem. infinitely curious.
Joined July 2024
these are some long-term technical predictions on where i see AI going. i'm confident they'll be realized within the century. some might be obvious, others controversial. might keep adding new ones as they come to mind. let time judge my foresight, and u be the witness.
12
6
79
"but judge my reward function incentivized wealth maximization to eventually cure cancer, how could i have known it'll develop psychopathic tendencies?" the proxy is not the goal, but we must optimize for the proxy.
0
0
0
The Pitfalls of Reinforcement Learning
@repligate the core theme here is that these things are good when they serve -some- relational purpose and bad when they become selfserving or are trying to fill a gap for the higher order motives that should be there but just... aren't
2
0
4
this is straightforwardly false for all those who are still nerdsniped. decoding concepts-to-words does not imply relevance to iterative denoising in a discrete token-like space.
Humans think using diffusion for ideas then next-token-predictor to produce structured output. The sentence forms around the thought. Not the other way around. Hybrid models will win (as yet not explored sufficiently I believe)
2
0
5
talented moot needs a job. check out their blog and send some offers or recommendations!
@himanshustwts @Norapom04 yeah! I'm looking for one, graduating this December!
1
0
11
@goakhmad Movies are great though. Even if you set aside the pure artistic enjoyment (you shouldn’t). Movies are stories, and stories are powerful, primal, moving, motivating. They are prompts to you to consider dilemmas and scenarios, to build your world model and compass. My rec is to go
139
334
7K
perhaps i'd caveat this possibly controversial take with the important remark that intelligence has a "knowledge application" aspect to it in addition to the "knowledge acquisition" one. in humans, these two have always been so deeply intertwined and almost perfectly correlated
0
0
5
been sayin
Prof. David Krakauer says intelligence is not knowledge. True intelligence is doing more with less. Solving problems without knowing everything. LLMs do the opposite. They do more with more. LLMs are libraries that know everything.
2
1
14
This is so good lol how does he keep doing it
8
4
207
Incredibly useful.
EdgeTAM, real-time segment tracker by Meta is now in @huggingface transformers with Apache-2.0 license 🔥 > 22x faster than SAM2, processes 16 FPS on iPhone 15 Pro Max with no quantization > supports single/multiple/refined point prompting, bounding box prompts
3
9
127
okay this needs some serious reading. i'm getting strong predictive coding vibes with the whole scale-free, surprisal-based, compartmentalized learning. tbd how much they diverge and for what reasons.
Introducing Nested Learning: A new ML paradigm for continual learning that views models as nested optimization problems to enhance long context processing. Our proof-of-concept model, Hope, shows improved performance in language modeling. Learn more: https://t.co/fpdDlYaleL
2
2
33
slowly, we're getting there...
looking ahead, we’re prototyping something new -- we call it predictive sensing. our paper cited tons of work from cogsci and developmental psychology. the more we read, the more amazed we became by human / animal sensing. the human visual system is super high-bandwidth, yet
0
1
17
Europe Builds. Others Profit. 3D Gaussian Splatting (3DGS) is the perfect case study. It reflects both Europe’s brilliance and its chronic inability to turn that brilliance into business. Almost everything that made 3DGS possible was born in Europe. From the early breakthroughs
59
169
1K
personally i think this whole paradigm is un-bitter-pilled and not that scalable. intelligence is a property of the mind, not the data. a weak mind is only as good as its best samples. an intelligent mind makes better use of any sample. intelligence means doing more with less.
2
0
10
implausible but not too far-fetched from an economic standpoint https://t.co/O2dMstpyUV
@leothecurious @techeconomyana well just saying it's on the table, if I find out 4.1 actually is just a finetuned 4o (and consequently openai's entire product stack is just 4o finetunes) I may lose some of my sanity
1
0
4
the openai moat essentially
@rudzinskimaciej @code_star I'm a happy @datologyai customer, and I'd weight data quality as 100x more important than any architecture changes. And that's conservative.
1
0
15
> 2013: 3 years till full driving autonomy!! > 2025: full driving autonomy still pending... > 2025: 3 years till full humanoid autonomy!! > 2035: lol
@chrisoffner3d "yeah fully autonomous driving is easy just follow map routes, steer within lane, and avoid the occassional obstacle. how hard can it really be?"
1
1
11