pama

10 ms·

181.
by pama 1y ago
This is as cool as it gets for using organic chemistry to design materials: design your own little lego blocks and let them self assemble into a humongous structure.
182.
by pama 1y ago
No it was never $100 for ChatGPT Pro.
183.
by pama 1y ago
Although I totally agree with this analysis, I also feel optimistic that this moment in time provides the first real opportunity in over 40 years for smaller organizations to start to affect societal change again. The existing efficiencies
184.
by pama 1y ago
I agree. A partial counterexample is the RL training loop on verifiable tasks, which uses the model in a loop to generate training data. Another one is the cleanup/prioritization of the pretraining data using earlier models. More gener
185.
by pama 1y ago
Out of curiosity, and IANAL, what is it in a GPL/copyleft license that would make it undesirable to train LLMs on projects with such license? Or are there stronger yet copyleft licenses that you had in mind? (FWIW, and not directly rel
186.
by pama 1y ago
> Viruses are unnecessary for prions (near as we can tell), and a nasty virus can do plenty of direct damage. Agreed > Someone could (if they discovered one) construct a nasty prion directly eh? And we’d probably be pretty fucked. It
187.
by pama 1y ago
If that was a genuine question, yes I have a very decent idea what prions are. We were discussing if the computational design tech should be giving people pause and the answer is no. If anyone really wants to create nasty stuff they dont ne
188.
by pama 1y ago
Yes it is technically a solved problem. But no you dont need to try to add a novel protein if you had nefarious intent. There exist more than enough well studied and fully understood examples.
189.
by pama 1y ago
Not sure what you mean. One can easily just make the existing prions if they wanted; they are good enough to destroy humanity if one finds a potent delivery mechanism. Again, the problem is not the protein printing tech or the protein desig
190.
by pama 1y ago
(Nearly) not concerned at all. “Printing proteins” is a routine procedure since the late 70s and it included novel proteins or variations of existing proteins since the early days of recombinant synthesis. Insulin and growth hormones were
191.
by pama 1y ago
Is there a link somewhere to the actual scientific contribution? (Or at leadt an abstract?)
192.
by pama 1y ago
Shingles comes to you after you are 70 in most cases. Please get the shingles vaccine well before then.
193.
by pama 1y ago
Having the local LLM process sensitive data is a desirable usecase and more trustworthy than using a “trusted partner” [0]. As long as your LLM tooling does not exit your own premises, you can be technically safe. But yes, dont then click a
194.
by pama 1y ago
Of course theoretically lots of things are possible with probabilistic systems. There is no difference with open source, openweight, chinese, french or american llms. You dont give unfettered web access to any models (locally served or oth
195.
by pama 1y ago
Why not chinese models served in house for western companies? Aren’t those what 80% of az16 startups use?
196.
by pama 1y ago
Thanks for the reminder to schedule the annual dermatology appointment.
197.
by pama 1y ago
There is room for improvement in inference, hence the presence of various startups in this space and the increased innovation in software. Large nvidia clusters are still cost optimal for scaling inference (as they move most of the memory
198.
by pama 1y ago
Although they are incredibly useful, the ML applications, or MLops, or ML Devops, or whatever other production-related application tech stack terminology may come to mind, are providing critical scaffolding infrastructure and glue but are n
199.
by pama 1y ago
Certainly all of humanity could approximate the reading of an LLM for now. The energy consumption of humanity is about 20TW at this point, so over 20,000 times higher than the next generation LLM training runs. Just like the LLM training
200.
by pama 1y ago
Sauro, if you read this, please refrain from such low-content speculative statements: “On a loose but telling note, this is still three decades short of the number of neural connections in the human brain, 1015, and yet they consume some on
201.
by pama 1y ago
Of course. Lots of useful metrics exist to help tweak code performance without always needing to go into detailed profiler traces. GPU utilization is a particularly poor metric in helping much, except for making sure the code made it to t
202.
by pama 1y ago
Wait until you encounter GPU utilization. You could have two codes listing 100% utilization and have well over 100x performance difference from each other. The name of these metrics creates natural assumptions that are just wrong. Luckily
203.
by pama 1y ago
If you make careful calculations and estimate the theoretical margins for inference only of most of the big open models on openrouter, the margins are typically crazy high if the openrouter providers served at scale (north of 800% for most
204.
by pama 1y ago
Agree that the writeup is very wrong, especially for the output tokens. Here is how anyone with enough money to allocate a small cluster of powerful GPUs can decode huge models at scale, since nearly 4 months ago, with costs of 0.2 USD&#x2F
205.
by pama 1y ago
Who is “we” (in the “We are launching”) and what is the mechanism you can convince people/investors that you or the proposer of the project know how to train large scale models and can make accurate practical estimates?
206.
by pama 1y ago
I agree. Your estimate for (2), about 0.0022 kWh, corresponds to about a sixth of the charge of an iPhone 15 pro and would take longer than ten minutes on the phone, even at max power draw. It feels about right for the amount of energy&#x2
207.
by pama 1y ago
Be careful when reading such blogs: > Note that the dosage in the mouse experiments is quite high — 0.1% of the body weight every day, meaning about 2 ounces a day for me (70 kg). Mouse and human metabolism are very different. A better s
208.
by pama 1y ago
OK. The more precise statement is that when the test was created you would have had about 99.3% to 99.42% chance that such a sample did not exist if you were to test all the population this test was designed for (depending on population sta
209.
by pama 1y ago
IQ of 200 (or higher) do not exist according to the original definitions of this metric. You need a population of 219 billion or higher to have a 95% chance that a sample exists with 6.66 standard deviations away from the mean (assuming mea
210.
by pama 1y ago
What was that thread? Oxford philosophy exams never suggest a particular stance on a question so you are free to argue on either direction of this argument and bring about your strongest points based on the current philosophical thinking o

More ›