Michael Bommarito
@mjbommar
Followers
1K
Following
3K
Media
535
Statuses
4K
RT @LSCtweets: Great news! The Senate Appropriations Committee just approved increased funding for LSC—$566M for FY 2025, a $6M boost from….
0
8
0
you need to spend more time watching C-SPAN2 and less time reading tpot.
The AI challenge is a moral one as much as anything. Are we going to protect individual rights and the working person, or are we going to allow the big corporations to run roughshod over the little guy and take whatever they want? We need to protect We the People
0
0
3
who cares about "big model smell" when you can research "big snack smell"
0
0
3
mark my words - a bipartisan coalition of strange bedfellows will make this is a first-order campaign topic into the midterms next year. if you talk to real people or pay attention to polling, you should not be surprised.
The Senate Judiciary Committee is gearing up to grill Big Tech. About time. These companies, and their AI arms, didn’t innovate. They stole. That’s not progress. It’s piracy. They knew it was illegal, and they did it anyway. Why? Because they thought they were untouchable. 17.
0
0
2
we've been mostly focused on allowing "larger" tokens in domain-specific applications, but it's interesting to see work that goes the opposite direction. lots of questions re: token->embedding, but the task results with ensemble are all that really matter in some sense (ignoring
Do you ever wish all LLMs used the same tokenizer?🧑🤝🧑. We present an *efficient, lossless* method to convert any LM into a byte-level model at inference time. This fixes weird tokenization artifacts at the prompt boundary and enables ensembles of LMs with mismatched tokenizers! 🧵
0
2
1
creating god-simulacra from warped versions of ourselves is cool and all, but have you ever tried simply enjoying the proverbial garden before the fall?
0
0
0
Big Dog doesn't want you to know that you can be your own hunting dog. you are a bipedal apex predator. literally just start running at the raccoon to tree it.
0
0
1
in one of our for-profit co's, we spent ~11k on opus from may to june before switching to $200/mo max plan for 3 people. .
Premium newsletter: Based on user data and their gross profit margins, I believe that Anthropic's most popular product, Claude Code, is haemorraging money on every user, with some costing the company hundreds - even thousands - of dollars in a few days.
0
0
1
been coding for the better part of three decades now. started using copilot nearly four years ago. and i can say, without a doubt, that coding has never felt more frustrating and less fulfilling than it does today. like the micro-managing the worst intern, but that's it -
1
0
1
i'm not saying we "invented" the same thing moonlight did, but it was def clear that a few tweaks to muon could go along way. we got perfectly stable runs with an even simpler solution in our kl3m-embedding models that was parallelization friendly: just adopt an adaptive
0
0
0
since og stable diffusion 3 years ago, i've struggled to find joy in photography - something i'd been doing for decades. i have rolls of undeveloped film just sitting in the fridge. lightroom folders never reviewed. just frozen in time. this shot was from august 2022, one of
2
0
2
two-fer test: big model smell and sino-sensitivity in one. you can tell who is a T and who is B. for no reason at all.
0
0
0
creative commons - an organization representing artists that abandoned X but loves linkedin - all but confirms their thinking. copyright and licensing are not their preferred policy tools, and therefore the tens of millions of people who rely on their existing license/deed
6
3
15
these latest tabular foundation models don't get enough attention.
👨🎓🧾✨#icml2025 Paper: TabICL, A Tabular Foundation Model for In-Context Learning on Large Data.With @JingangQu, @DHolzmueller and @MarineLeMorvan . TL;DR: a well-designed architecture and pretraining gives best tabular learner, and more scalable.1/9
0
0
2
we don't have rocky mountain up here (yet), but damn. the deer population is rapidly becoming a timebomb for deployment of pathogens that would make nation-states jealous.
0
0
1
i'm deeply concerned that CC's leadership has been completely captured/aligned to big tech thinking (or money). it's either that, or so much pure hatred towards the concept of copyright that they would rather let the commons be destroyed than possibly find any kernel of good in.
0
0
1
maybe it's the days of fever > 104 talking, but after reading the CC signals project documents, i'm going to call it - this is textbook bullshit. the plan is, in their own words, based on adhesion to social contract. they expect big tech to follow norms because social pressure.
1
0
0