theseriousadult
@gallabytes
Followers
7K
Following
214K
Media
185
Statuses
4K
father, ML enjoyer. @midjourney v2-7.
Oakland
Joined April 2014
everyone is building really big short-term memories and calling it long-term memory. they're different things. almost nobody is building long-term memory. I barely ever see papers on it. it's the last remaining puzzle piece imo. should be the thing people are trying to build.
29
14
324
gpt5 really loves \boxed answers and the chatgpt mobile app refuses to render them. much better in the web app.
1
1
11
I have seen people say this but this literally isn't a 3d model? I guess you could make multiview and hope it's consistent enough to run through a real multiview -> mesh pipeline?.
Nanobanana from Google just dropped today. The best use case I heard so far is taking objects out of pictures and creating 3D models from them for games!. Any THING from a movie can be put into a game.
7
0
22
google people have started using x like openai people. kinda cringe. not a fan.
15
3
252
and, similarly to how claude code doesn't assume you're a good typist, a good image generator assumes you're a good artist but not necessarily especially skilled with a pencil.
Claude Code works because it assume you know how to code- it lives in the terminal. I wish more AI tools would do the same. Image gen tools that assume you're a good artist, writing tools that assume you're good at writing. Not just prompt -> output.
0
0
6
opus 4.1 seems qualitatively better in research mode than 4.0. seems like it's gotten a lot less long winded, does more synthesis and summary less pure restating of source content.
2
0
13
throwback.
1
0
18
filed under "reasons you can't just train an out of date model and make up for it with tool use". you wouldn't expect to need to call the search tool to know something as common knowledge as "who is president"!.
3
0
17
RT @gallabytes: I feel like there's an interesting @patio11 style piece that could be written on the regulatory constraints of investing OP….
0
1
0
I feel like there's an interesting @patio11 style piece that could be written on the regulatory constraints of investing OPM and why the reader's oh-so-clever simple tweaks to common practice might be less adaptive than they think.
1
1
3
my hot take is that investors are underexcited about AI and overexcited about "AI" and this is basically downstream of the same regulatory barriers that create most of the other toxic vc dynamics.
be Sam Altman.openai gonna spend TRILLIONS on datacenters soon. We have better models and no, you can’t see them lmao . “Are we in an AI bubble?”.Altman: My opinion is YES.
2
1
21
the only posts I see about gork imagine are people dunking on it. none of the posts they're dunking on make it through my feed's slop filter. your feed can be good you just have to want it hard enough.
2
0
10
someone just tried to spearphish me to take over this account. reminder to everyone that this site doesn't have a phone support line and nobody ever calls you to request a confirmation code. also if you're the guy who called me and you see this post, hi! good try lmao.
4
1
25
new Claude conversation search feature is nice. glad to have it, glad it's an explicit tool call vs hidden, wish it also had a memory feature to write things down which Claude knows I can't write to, possibly tagged by which model wrote the note.
0
0
6
this is a really cool approach and I'm frankly surprised it works!.
Bifrost-1: Bridging Multimodal LLMs and Diffusion Models with Patch-level CLIP Latents. "We present BIFROST-1, a unified framework that bridges pretrained multimodal LLMs (MLLMs) and diffusion models using patch-level CLIP image embeddings as latent variables, which are natively
0
0
7
long term memory posting will continue until the models can sleep.
the reason why gpt-5 can't do <insert thing humans can do> is that we've pretty much exhausted what's possible using a single-turn chat format. Yes, chatgpt can't discover new science, but neither can humans in ~20 minutes with only access to a search API and a python notebook.
3
2
56
"o3 with way fewer hallucinations" is actually a very good model concept and I am glad to be able to use it. I am still a bit skeptical of the small model plus search instead of big model with big latent knowledge style, but within those constraints this is a very good model.
0
0
25
gpt5 thinking is good but way too slow even for easy things. gpt5 not thinking is not very good. need gpt5-thinking-low.
2
0
20
wish image generators could be tools for communication like this.
visual forms hold the power not just to represent knowledge, but to create it. [Athanasius Kircher's diagrams]
2
1
38
💯.
Language models such as GPT-5 are missing a Midjourney-esque wonder type quality and i will not stop talking about it.
2
0
19