iiinova — where your digital self keeps evolving

crumb is for techies

050

we have tons of inference-only kernels that are INSANELY fast, hundreds to thousands of words per second. the training kernels can't do that speed yet. BUT if we expose token embeds as an input method, we could just treat the sys prompts as embeds to optimize a la prompt-tuning...

010

train the prompt yeah, just get the encodings of tokens for a seed prompt (or, i suppose you can start from random) and pass them to your optimizer. backprop through the model to the encodings. though i think evolutionary methods might be able to use this though because—i have to split this, cont-

110

asimov 1991, watch in full: www.youtube.com/watch?v=gGib...

000

crumb @crumb.bsky.social · 23h

back in ye old times google found you can finetune the meanings of words in a prompt and it sometimes will outperform even full model finetuning. also, it wont break special megakernels like lora does :) www.geeksforgeeks.org/artificial-i... arxiv.org/pdf/2104.08691

Figure 1: Standard model tuning of T5 achieves strong
performance, but requires storing separate copies of the
model for each end task. Our prompt tuning of T5
matches the quality of model tuning as size increases,
while enabling the reuse of a single frozen model for
all tasks. Our approach significantly outperforms few-
shot prompt design using GPT-3. We show mean and
standard deviation across 3 runs for tuning methods.

Table 1: F1 mean and stddev for models trained on SQuAD and evaluated on out-of-domain datasets from the MRQA 2019 shared task. Prompt tuning tends to give stronger zero-shot performance than model tuning, especially on datasets with large domain shifts like TextbookQA

120

i just realized like ~nobody here was around for prompt tuning because it was 2021 🤦‍♀️

011

crumb @crumb.bsky.social · 09/10/2026

or i just want an excuse to send big unreadable files to people back and forth again

000

crumb @crumb.bsky.social · 09/10/2026

we are prob mega under utilizing finetuning tokens a la prompt tuning & to less extent those vram maxing dreambooth configs. i wonder if good inference framework accepting inputs_ids would open some things up. evolutionary methods prob work stellar there

230

crumb @crumb.bsky.social · 08/10/2026

please let this project be a somethingburger

140

Reposted by crumb

Colin @colin-fraser.net · 08/10/2026

Calculators are impressive

317912

Doll @dollspace.gay · 07/10/2026

sorry guys, magic is real

060

Reposted by crumb

What do the moderates want? Solar panels at the data centers? Sounds fucking great. Water offsets? We will help you push for that. Bans on deepfakes? Awesome we dont like that shit either.

11568

have a blissful singularity festival

030

maybe openai can solve my parents next

1104

i dont think theres a bottom

220

im supposed to go to the dentist that's what im supposed to do

140

yeah i think it's just gonna happen here throughout the next decade. idk what im supposed to do rn

140

my uncle was on the phone literally two days ago saying "ai is all just smoke and mirrors, it can only say things people have said before"

1120

sorry I don't really have anything substantial to say about the math thing other than all computers will be able to do this soon for all other fields of study

3502

my gpus arent going to run today im taking an outside day

030

all i can do is drink my smoothie pretend everything is normal today

131

scanstone.bsky.social @scanstone.bsky.social · 15/09/2026

news.ycombinator.com/item?id=4998...

ihoggan
I was a graph theory junkie long ago and even moved to Budapest for awhile to study among the greats. While I was there I started working on Barnette's Conjecture which came to occupy my thoughts over the next 24 years of my life, on and off as I worked in many different fields. Last summer I even thought for a few days that I had actually solved it. But it's supposedly proven here - problem 180. I don't know what to think exactly. I spent thousands of hours on that problem. I really enjoyed it. Hearing that it is solved somehow makes me sad in a far-off way, like hearing an ex-girlfriend died suddenly in a car crash. I don't know, there's probably a lot of people feeling odd emotions tonight. There's no Lean proof for this one so I'm digesting the paper. On the surface it looks like an approach I considered 24 years ago and abandoned.
I revisited the problem this summer, along with my partial solutions, when the previous round of stunning proofs came out. Several hours of work with Fable simply convinced me it wasn't yet solvable and reinforced how hard of a problem it was.

jboggan
I went back to that Fable chat and showed it this new preprint. It coded up the new constructive algorithm and ran it against the existing test suite, that looks good at least.
It has been super helpful in delineating where the crucial concept came from. The proof is rather simple as graph theory proofs go, but it does seem to use some constructions that would only seem obvious if you had serious physics experience with partition function and calculating energy states that cancel out. It's not a wholly alien bolt from the heavens, but I can also see how there hasn't been a human being with the broad theoretical physics knowledge combined with the deep graph theory experience in planar graphs to come up with this idea. I don't know, I'm looking for precedents of this formulation and some old papers of Penrose counting the number of edge colorings of this same graph type are coming up, the line of argument at least rhymes. But I agree with the thought that this sort of progress should not be siloed inside those companies. I propose a tax so that every slop cannon Al video pays for another hour of compute time for advancing mathematics.
reply

823452

Reposted by crumb

I'm a published mathematician - I worked on the question of whether the unit balls of Banach spaces are plastic. OpenAI's GPT-6 Astra is *far beyond my abilities* in this question. I have read everything written on it. THERE IS NOTHING TO PLAGIARISE, I READ IT ALL AND IT'S NOT THERE

2854

wow i need to create a Highly Persistent Model of my own

030

brainfuck swarm from zero needs an insane amount of compute that i do not have the time to afford it. will rethink main algo while i do other things

010

new era of software is ACTUALLY here yall

091

actual computer just released toks, "the best tokenizer on earth and the first tokenizer designed for the quadrillion-token era." most of the work is done in pure assembly. genuinely nuts. actual.inc/company/blog...

39715

artificially induced gaia is my superpersuasion

040

the field is not "bury agents in systems" sorts of complex yet but it does point the way, i think

120

search: "physical reservoir computing"

130

an asi burying agents to do its bidding in dynamic processes like the weather or social interaction (or, really, a distributed process across many many weakly interacting systems...) is, i think, the most yud thing that i imagine regularly

170

im imagining im just missing the whole reading bit, electrodes + board they get connected to before going to microcontroller? i have not dug too deep yet

250

this would bring it much closer to feeling like an "exocortex" than anything else i could do IMO

0120

i like. Really want to have a subvocalization flow like "uplink. {prompt}. [silence]" -> send to llm agent on my workstation over wifi -> earpiece. i want this more than a lot of things and i think i can do it

2141

i just need to find like $25 for the hardware bits i dont have

1121