HN user
http://www.moderndescartes.com
(author here)
As a serial successful field-hopper, I agree that I'm not the right person to be making these estimates.
But the external view is that college courses roughly expect you to do what I'm claiming, in roughly the time investment that I'm claiming -- and undergrads are typically in 4+ classes at a time. So is it that the whole educational system is delusional? (I fully acknowledge it might be so!)
I could have sworn that sickly sweet smell was the smell of various phosphine reagents? Just my vague recollection of my time in lab from 15 years ago.
Does this handle, e.g., water of hydration CaSO4 . 2H2O? states of matter H2O(g)? does it preserve subunit information, as in (C6H5)CH2COOH? Writing a parser for basic formulae is such a tiny tiny part of the actual problem... deciding the scope of what you want to handle and how is the real problem
Great write-up, but I admit that I found the interweaving of human and AI-written content/headlines/summaries pretty distracting. I kept on wanting to scroll past, but had to keep on backtracking to find the human thread again.
I think if you want to give your reader a quick intro to, e.g., what is the Adam optimizer, a simple link to Wikipedia is fine. No need to copy-paste an AI tutorial on Adam into the blog post.
AI slop
Taste is the field behind the goalposts, and we as humans are constantly expanding that field by "having taste" and developing taste as a community.
For those commenting on cost per token:
This throughput assumes 100% utilizations. A bunch of things raise the cost at scale:
- There are no on-demand GPUs at this scale. You have to rent them for multi-year contracts. So you have to lock in some number of GPUs for your maximum throughput (or some sufficiently high percentile), not your average throughput. Your peak throughput at west coast business hours is probably 2-3x higher than the throughput at tail hours (east coast morning, west coast evenings)
- GPUs are often regionally locked due to data processing issues + latency issues. Thus, it's difficult to utilize these GPUs overnight because Asia doesn't want their data sent to the US and the US doesn't want their data sent to Asia.
These two factors mean that GPU utilization comes in at 10-20%. Now, if you're a massive company that spends a lot of money on training new models, you could conceivably slot in RL inference or model training to happen in these off-peak hours, maximizing utilization.
But for those companies purely specializing in inference, I would _not_ assume that these 90% margins are real. I would guess that even when it seems "10x cheaper", you're only seeing margins of 50%.
Desalinated water is also less dense than normal seawater, so the water column inside the output pipe would create a pressure imbalance with the water column outside the pipe, assisting in the outflow? I'm having trouble figuring out how to resolve this seeming perpetual motion machine
I'm working on a startup in this space and wrote up my thoughts here: https://www.moderndescartes.com/essays/study_mode/
As a former big tech engineer, I can't help but come up with a gazillion ways to work around these sorts of seemingly straightforward policies.
Here's one way they could get around their own privacy policy: keep track of what % of Claude-generated code is retained in the codebase over time (as an indicator of how high-quality / bug-free the code was); A/B test variations of Claude Code to see which variations have higher retention percentages.
No usage data is retained, no code is retained, no data is used (other than a single floating point number) and yet they get to improve their product atop your usage patterns.
Here's another idea: use a summarization model to transform your session transcript into a set of bits saying "user was satisfied/dissatisfied with this conversation", "user indicated that claude was doing something dangerous", "user indicated that claude was doing something overly complicated / too simple", "user interrupted claude", "user indicated claude should remember something in CLAUDE.md", etc. etc. and then train on these auxiliary signals, without ever seeing the original code or usage data.
Back when Claude Code had per-token pricing, almost nobody used it because it was clearly much more expensive than the Cursor pricing - $20 a month flat for Cursor vs $5-10 a day for per-token Claude. The incentives manifested in the way both products used tokens - Claude Code has no particular qualms about submitting a gigantic number of tokens and letting Sonnet figure it all out, whereas Cursor puts in a lot of traditional software engineering to figure out the correct minimal context to put in. Now that Claude Code is on a fixed price plan, it strangely doesn't seem like Anthropic is doing anything to optimize the number of tokens consumed by Claude.
I think it's quite plausible that Anthropic is bleeding out ~100/month on token costs per $20/month user, and even at 80% margin, this is just merely breakeven. Their limited capacity also means that they are _losing_ the opportunity to sell the same capacity at a per-token marginal profit. I think the only plausible endgame here is that Anthropic uses the usage data to RL-finetune Claude Code to the point where it is actually worth a $200/month subscription.
Enjoy the $20/month Claude Pro plan while it lasts; I don't really see it sticking around for more than a year at best.
I'm a second-gen Korean-American; my korean is weak but conversational. I am intrigued by the reasoning model that analyzes my speech and points out various mistakes I'm making. It's a good first attempt at separating the 2 tracks of actual conversation vs mistake-correcting.
I think showing the raw reasoning text is not quite the right UI; maybe highlighting the specific text in red and showing a suggested correction would work better?
It's also a little awkward that the conversation is live; I don't really have any breathing room to read the reasoning traces on what mistakes I made / could have done better. I hung up the first time I tried to figure out how to pause.
For those of you suggesting we use the extra energy for $pet_topic I suggest reading https://www.moderndescartes.com/essays/factobattery/
Tldr: most applications of free energy have capital costs that far outweigh the free energy harvest potential.
The loop can be anywhere, not just the north pole :)
I am building such an AI tutoring experience, focusing on a Socratic style with product support for forking conversations onto tangents. Happy to add you to the waitlist, will probably publish an MVP in a few weeks.
Followed an identical path.
See my source code here:
https://github.com/brilee/modern-descartes-v2/blob/master/ma...
Includes:
1. RSS feed
2. Blog listing pages ordered by date
3. Tagging system
4. Localhost dev server with file-watching recompilation step.
https://www.moderndescartes.com/essays/gnugo_to_agz/
I happen to have recently written up a longer history of Go AI. If you're wondering about what is special about Go in particular or what generalizes to other problems, give it a read.
Pretty cool result with lots of experimental evidence.
Another purely speculative example - Geosmin is the human-recognizable indicator of rain, and given how intensely sensitive we are to this molecule, I wouldn't be surprised if many ecosystem participants also sniff for this molecule.
Oh! I think this explains the issue I had on Android Firefox where every so often my phone would randomly consume 10gb/day in data (costing me $stupid on phone bill). I believe a news site was playing audio and was loading new ads on repeat and the tab never went to sleep because of the audio! I had to switch to chrome because the repeating surprise data bills were becoming silly
This is a really belated blogspam repost. Original:
Being one of those lucky few at Google Brain who switched into ML early enough to catch the wave... I fielded my fair share of questions from academic friends about how they could get one of those sweet gigs paying $$$ where they could do whatever they wanted. I wrote these two essays basically to answer this class of questions.
Why did Brain Exist? https://www.moderndescartes.com/essays/why_brain/ Who pays you? And why? https://www.moderndescartes.com/essays/who_pays_you/
Free lunch briefly existed for a small lucky few in 2017-2021, but today there is definitely no more free lunch.
Why does this response sound llm generated to me? Maybe it's the phrase "however it's worth noting"
His argument seems to boil down to a the financial regulator equivalent of the trolley problem.
In Operation Choke Point, the regulators were actively moving the lever to debank morally repentant industries. In "Operation Choke Point 2.0", the complex system of regulatory guidelines and actors seemingly self-coordinates to debank crypto, and the failure of anyone to intercede is painted as a willful neglect. Those on the regulator side say, "I didn't do anything", and those on the crypto side say, "you have every power to do something".
Choral/string music is quite nice - the musicians have an opportunity to bend pitch as locally appropriate to make perfect intervals. I definitely did this when I was singing in a chorus. But piano is still my first love :)
Perfect pitch means that I can sit down at a piano (or other instrument), play one random note, and you can instantly tell me what note that is. No preamble, no tuning yourself up, you can just do it.
Yes, I can do this. I don't understand why people are so keen on gatekeeping perfect pitch.
It sounds like you had something similar to what I had with the clarinet in middle school.
added!
editor's notes:
Very, very obviously a Chopin Waltz. The chord choices, usage of triplets/mordants, and the suspended pedal tones is characteristic. It's a mix of [Prelude Op 28 no 11](https://www.youtube.com/watch?v=si5aT6FDPZ0)'s ephemeral brilliance and [Waltz Op 34 no. 2](https://www.youtube.com/watch?v=PGdpRmL2XUc)'s lilting, moody style.
My "edition" is definitely not an [urtext](https://en.wikipedia.org/wiki/Urtext_edition). I made minor simplifications to the descending line at measure 8, following Lang-Lang's performance (which I believe is the correct decision - the arpeggiated diminished chord F-D-B-G#-F-D sounds better without any gaps, compared to what's in the scan, B-F-B-G#-F-D.), and added some phrasing where I thought it was obvious and perhaps went missing over the ages from the raw scan. The ornamentation in measure 20 was probably modified by Lang-Lang, but I think it fits the piece better than a plain mordant, so I notated it as played.
If I were to judge based purely on the music - minus all of the contextual clues like paper, ink, backstory - the probability of it being fake is ~10%. I say this only because it is shockingly similar to 34-2 in harmonic and stylistic elements, which is exactly the kind of thing an AI trained on a not-big-enough dataset would do. While AI utterly fails at longer pieces, it could plausibly render a coherent 24-measure piece in the style of Chopin, and DeepMind could plausibly be working in stealth on a really good music-composition AI. But in the end, the piece is too tightly composed, and I trust NYT's decision to trust the historians who are familiar with evaluating such artifacts.
I transcribed the score, available here:
https://www.moderndescartes.com/essays/chopin_waltz_posthumo...
fp4/fp8 for neural networks don't work the way you think they do - they are merely compression formats - a set of, say, 256 fp32 weights from 1 neuron are lossily turned into 1 max value (stored in fp32 precision) and 256 fp4/fp8 numbers. Those compressed numbers are multiplied by the fp32 number at runtime to restore the original weights and full fp32 multiplication + additions are executed.