james @http_413

@landontgreen nobody's actually buying these from amazon unless they're poorly informed or money is no object. they're lingering around $8.5k and up but nobody's biting. meanwhile they're back in store at micro center at $6,950 and not flying off the shelves like they were.

1 0 2 313 0

View Details

james @http_413

a day ago

@LLMJunky but can spark run on a spark?

0 0 1 14 0

View Details

james @http_413

a day ago

@MiaAI_lab x.com/http_413/statu…

james @http_413

2 days ago

@Ananth7e google doesn't get that ai hype cycles turn so quickly that they already lost their chance to release. gone are the days when you could tease a model and string people along for weeks or months. nobody is going to care if/when it finally does come out. people have moved on.

0 0 0 100 0

0 0 0 23 0

View Details

james @http_413

a day ago

@collision @finkd @alexandr_wang drinkin these beers

0 0 0 67 0

View Details

james @http_413

a day ago

@mark_k @imh_habib memini?

1 0 0 43 0

View Details

james @http_413

a day ago

@Ananth7e don't know why people always jump on the company's released benchmarks instead of waiting for real world testing. how many times is this going to happen before people learn?

2 0 4 99 0

View Details

james @http_413

a day ago

@Polymarket should've deployed le chonk for defense

0 0 0 13 0

View Details

james @http_413

a day ago

@MiaAI_lab tried nemotron with tensorfold but the stream just kept dying in the middle of responses.

0 0 0 100 0

View Details

Noah Giansiracusa @ProfNoahGian

3 days ago

When Unit Distance conjecture dropped I was dismissive: this is just one combinatorial problem that benefitted from a heavily computational approach. When Navier-Stokes dropped I was deeply rattled and uncertain: this is a huge problem, but not a problem ordinary mathematicians work on and it cost millions of compute, so not the budget ordinary mathematicians have. Clearly a huge development but felt hard to disentangle what it really portends. When the nearly 400 problems recently dropped, including ones I had used for years in grant proposals as North Stars guiding my research, I felt… relieved and optimist! I didn’t expect to, but I realized that this takes the uncertainty and anxiety away—it was a deep cut of problems across math, each solved on average with only 3hrs of complete, from a model that will soon be in user hands. The writing on the wall is unmissable now: math is entering a new era. Not just millennium problem math, but ALL math research. I find this exciting and feel more optimistic than I have been until now— because we can stop fretting over how capable AI is and how serious the threat to the profession is and move on to acceptance and finding ways to make the most of this powerful new tool—which I think will be thrilling! Just as one tiny example of an exciting prospect that comes to mind. OpenAI solved about 400 of 4000 problems—why did these fall while the others didn’t? Imagine defining a new difficulty measure, somewhat akin to Kolmogorov complexity, that’s roughly how hard a math problem is to solve with an AI (I know this is extremely ill-defined and not practical, but just humor me for a moment). As we solve some problems and fail at others a landscape of values of this parameter emerges. Among human-hard problems, what kind of range is there for this AI difficulty measure? Why? Which problems and fields have the highest value in this measure? Can we prove upper or lower bounds on it? What problems have the highest values? It will be amazing to look back at the millennium problems and see the wide range of AI difficulty they posed, and to try to understand why that is. This is just one random thought, but it’s an exciting avenue that will tell us so much about hard problems that we never could even ask before, a new window of research that can open now that AI in math is the new normal. There’s so much else that we’ll be able to do and explore in this human-AI hybrid math system, I’m finally feeling ready to embrace it and honestly quite excited about it!

47 57 765 53K 184

View Details

james @http_413

2 days ago

@brankopetric00 the old short-the-outlet-to-see-which-breaker-trips technique

1 0 6 751 1

View Details

james @http_413

2 days ago

@the_varunrana now my team's jira tickets are 5 page essays via llm

1 0 15 311 0

View Details

Brian Armstrong @brian_armstrong

2 days ago

We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case). I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves. We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them: "/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started" I imagine we'll have some backend services that could be replaced entirely by this new architecture.