Maciej Fijałkowski

@jiayi_pirate What are the minimum RAM/VRAM/GPU requirements? I keep running out of nvram in trying to reproduce (at ~80G)

One caveat, of course, is that it's validated only in the Countdown task but not the general reasoning domain. We are now bounded by compute, and please reach out if you wanna help!

We reproduced DeepSeek R1-Zero in the CountDown game, and it just works Through RL, the 3B base LM develops self-verification and search abilities all on its own You can experience the Ahah moment yourself for < $30 Code: github.com/Jiayi-Pan/Tiny… Here's what we learned 🧵

@Plinz According to this fred.stlouisfed.org/series/GDP (and others), the US GDP grew like 50% in those 4 years... which is hard to comprehend? I'm not sure I understand any of it, the growth is more like 2.5%, but any graph of GDP shows much faster growth

I get vertigo from statements like "The US economy is doing great!" and "The US economy is the envy of the world!". The US has been overdrafting its credit cards to buy candy like there is no tomorrow. How can we avoid a crash?

@samdknowlton does that mean you can look for mushrooms using infrared camera?

Mushrooms possess a remarkable and previously unrecognized mechanism to actively disperse their spores, challenging the long-held belief that spore dispersal is a passive process at the mercy of external winds. This mechanism relies on evaporative cooling and convective airflows, allowing mushrooms to exert control over spore dispersal even in low-wind environments or when crowded together. The process begins with the high rate of water loss from mushroom surfaces, which far exceeds that of plants. This evaporation cools both the mushroom and the surrounding air by several degrees Celsius. The cooled, dense air beneath the mushroom cap then spreads outward as a gravity current, carrying spores with it at speeds of several centimeters per second. Crucially, mushrooms generate asymmetric airflows beneath the cap to create effective dispersal. This asymmetry can be achieved through variations in the cap's shape or temperature gradients across its surface. The distance spores travel increases with the square of the gap height beneath the mushroom cap, allowing mushrooms to control dispersal by adjusting their morphology. These convective flows can transport spores from gaps as narrow as 1 cm and lift them over surrounding barriers. When spores encounter vertical surfaces, they can climb upwards, potentially dispersing even further. This mechanism explains how mushrooms can tolerate and even benefit from crowding or growing close to the ground. The high water needs of mushrooms, rather than being a disadvantage, are key to generating these dispersive airflows. This mechanism allows for effective spore dispersal in various environmental conditions, demonstrating how mushrooms have evolved to actively influence what was previously considered a passive phase of spore dispersal.

@fzzzy7 what is your weapon of choice?

spending a lot of time trying to make people feel uncomfortable listening to my music

@katclone What do you need? I kinda like making tools for people or optimizing their software to run faster, weird hobby

can we get some SWEs into biology labs on a consultant / retainer basis ???? most code written by biologists is so painful 😿

@fzzzy7 No, we must replace the human capabilities with AI-powered system with absolutely no feedback loop. Prepare for your slow death, while your limbs get corporate logos.

Englebart's vision of the augmented human and what a tight feebdack loop means was sidelined and forgotten. Augment yourself

@FlorentWaltz Hi Florent. Can you share with us 3D models of that stuff so we can look at it in Virtual Reality?

Super excited to share our latest work on the native in-cell organization of the mitochondrial respiratory chain 🥳 Using cryo-electron tomography🔬, we show how the respiratory complexes (and other complexes) are organized inside native mitochondria!  biorxiv.org/content/10.110…

It's an excellent time for both sides of the divide in America to admit their candidate and the selection process is at least somewhat flawed. It's incredible how many folks refuse to do that. Self awareness does not hurt yo!

@ibuildthecloud My python, my biology interests, my AI interests, my ....

this is how my python typically goes 1. git clone repo 2. create venv 3. run pip install 4. something fails 5. complain on twitter

@moxie Is the source code/prompts/whatever posted somewhere?

I made this last weekend to experiment w/ building an app end to end on LLMs: vibecheck.market It's like Wirecutter, but uses an LLM to recommend product choices based on reddit conversations and reviews, so you don't have to spend 20-30min reading reddit My experience: I'm late to the game, but this is the first time I've tried building an app end to end around an LLM. 1. It's very fast to build something that's 90% of a solution. The problem is that the last 10% of building something is usually the hard part which really matters, and with a black box at the center of the product, it feels much more difficult to me to nail that remaining 10%. With vibecheck, most of the time the results to my queries are great; some percentage of the time they aren't. Closing that gap with gen AI feels much more fickle to me than a normal engineering problem. It could be that I'm unfamiliar with it, but I also wonder if some classes of generative AI based products are just doomed to mediocrity as a result. 2. On the other hand, I think the discomfort I feel with having an unpredictable black box at the center of a product which can fail in very creative ways some percentage of the time might actually be a competitive advantage for startups. I think Wirecutter, especially branded as NY Times, would have a lot harder time tolerating that unpredictability and unreliability than an app like this (or an actual startup) which can set that product expectation with users to begin with. It might be that the current problems with generative AI are actually the things that create an innovator's dilemma and give startups an advantage to slip under the incumbents. It seems like the ideal spot right now would an app where 90% "done" is a mostly great experience, but the failings are still somehow not stomach-able by incumbents. 3. I do not understand how the economics of LLMs pencil out. When I look at the per concurrent user costs associated with inference, they seem orders of magnitude higher than per concurrent user costs of previous internet technologies. It seems to me that if previous apps like webmail, messengers, etc had costs as high, they would not have been viable products. This is something I want to learn more about.

@soychotic I got autodesk revit to land in a state where it would neither install nor uninstall, consuming gigs of space. I think I refuse to ever install anything from autodesk not in a VM, too sketchy

And sooooome programs (cough autodesk cough cough) leave behind gigabytes of nonsense that don’t even have anything to do wit me

I truly do not understand why applications leave their config and cache files on my computer after I uninstall them like an ex girlfriend “accidentally” leaving her scarf that she will inevitably use as an excuse to meet up again

@ID_AA_Carmack As someone running a business with google reviews, the "5 star is no problem" attitude means that 1 star reviews (even fake) yield very disproportionate impact on the overall rating. Averages are just removing too much information I suppose

I do conform to the expected Uber standard — a ride with no problems gets five stars. An exceptional experience gets an exceptional tip. I have been applying the centered ratings to comixology/kindle comics, and I want a “filter by 5 stars” option so I can find them again.

My “out of five stars” ratings are Gaussian, with most being three stars, and five stars being “standout good” and one star being “I resent having spent time on it”. This clashes with the standard for Uber and such, where five stars means “no problems”, and highlights the challenges of aggregating statistics on subjective measures. When a site says “1630 5 star ratings”, you expect that it is literally summing up things users clicked on, but there is more information available if the users have rated enough things to allow understanding their rating distribution. I’m sure this has been analyzed to death in recommendation work, but it is interesting to consider ELO like pairwise comparisons between products across users with different rating styles and comparing against simple summation. Of course, if you have sufficient users to find enough overlap in pairwise comparisons, simple averaging is probably already fine.

@Pinboard I think it should be a basic human right to eat exclusively peanut butter sandwiches free of scurvy!

I can't believe I forgot this. Yes, this is the same guy who giant-brained himself into getting scurvy twitter.com/NoHotSauce2Hot…

There is no IQ cap on stupid

@zeeg I mean look at us, you are trying to shit on an information acquisition strategy you don't understand and I'm here!

@zeeg Shitting on something gets people to do their best work defending. It's a lot like being wrong on the internet is so much better for information acquisition than asking a question.

I will never understand the game people play by trying to shit on things they don't understand. e.g. a bootstrapped company shitting on a venture company Be proud, but don't be ignorant!

@markwragg javascript

@Pinboard Meanwhile in South Africa

If you're a guy over 12 years old, not poolside or within a hundred meters of the beach, and you can feel sunlight hitting your knees, you need to go home and put some pants on.

Men's shorts should reach down to the midpoint of the shoe

Do our cells also feel like they're special and have destiny?

@paulg You get a similar graph (with less growth) at a climbing gym that's doing well, where all the growth happens in Jan, but then it stays

A novel variant of exponential revenue graph. This company is selling something users often sign up for as a new year's resolution. The product works, so they stay around.

There is a bit of uncertainty in the numbers, but it looks the projections for GDP growth are actually worse for UK than for Russia in 2024 and 2023. Sanctions much....