10,978 karma · joined January 30, 2013
roshan@arjie.com
[ my public key: https://keybase.io/roshangeorge; my proof: https://keybase.io/roshangeorge/sigs/6EQyu7jh7oa-KjkTAQBoJJ61fE2RBiZ546abgZc6zsM ]
Blog: https://wiki.roshangeorge.dev/w/Blog
I use https://overmod.org to manage highlight/block
For the same reason that we collectively believed wooden sail ships were a good thing to have and now almost none of our ships are made of wood or powered by sail.
Things changed. The top mathematical results of the last decade did not come from academia. It came from machine intelligence labs. In the last month.
Theorycrafting is all well and good but if you claim that some metric X (farebox recovery) is negatively correlated with some outcome Y (frequent good transit) then you do have to explain why all the good outcome Ys actually have high X. Reality has to meet the theory.
If what you said was true then it would be the inverse - operating costs would be mostly paid out of government subsidies in the areas with most ridership. But in fact, it is (and predictably so) the case that good transit systems have high farebox recovery - the HK MTR, Taipei MRT, and the Tube are those I’m familiar with and they’re all much nicer and more frequent and have high farebox recovery.
In fact, I find the whole thing confusing. Do people actually believe public transit is meant to be some kind of giveaway to a small number of riders? It’s clear from looking worldwide that where it is useful riders pay high enough fares because it is so useful.
I can understand some sort of bootstrap capital investment (or one-off interventions like for COVID-19) but the idea that operating costs should be predominantly covered from general taxation doesn’t really hold up if you look at which places are effective at this.
Skytrain bypasses this by not having drivers and perhaps as buses become AVs we could add many more of them here in the Bay Area with much higher frequency.
> That is why the model needs far more instruction to write than to code. I deploy it for its strengths and am constantly on alert for its weaknesses, which is what makes it a tool and not a ghostwriter.
Everyone is convinced of these grand truths - that they alone have contributed something of value to the machine work. But that value is extractable by almost anyone who interrogates the machine.
Like everyone else alive, I too believe this is monocausal. This is all a problem caused primarily by a distinction between the one who sees the supposed benefit of the academic (therefore induced to pay) and the one allocating resources to the academic. It is time we remove this intermediary who does this. Should you desire to read analytical writing by a professional machine speaker there is no doubt he will gladly take your money to talk to the AI for you. Let there be a market for academia - the mathematicians can be paid for the beauty they wish to glimpse between coffees and the philosophers can prompt their machines for money.
For the rest of us, release us from this compulsory taxation so that the plumber - should he so desire - can ask Fable 5.1 "Why should I pay academics to use AI to write?" and receive a satisfactory answer without needing to pay out of his taxes a compulsory fee for an answer to a question he may never desire to ask.
I don’t mind waiting but I think I’d prefer paying for faster results. BART took half a year to respond to how they calculate on-time rating but they did finally give me the answer. It had some interesting properties (canceling trains leads to higher on-time performance if trains are late) and I wrote it up here: https://wiki.roshangeorge.dev/w/BART_On-Time_Metric
Grading homework is a measure of conscientiousness and perhaps a pedagogical tool to force learning but by college perhaps we should have moved out of the need for the latter.
One can simply measure conscientiousness through attendance grade and mastery through exam grade and then we have two grades the universities can certify students for that are no longer dimension reduced into a single number.
Interestingly, the coordinating body that selects books, facilitates distribution, and negotiates wholesale pricing is the Dolly Parton Imagination Library, but per-book funding is actually entirely through local charities and so on. This means that where charities are unable to act as partners, children will not receive books, but it also means that the program can operate by growing and shrinking as local support (including from state and local government grants) comes online and goes offline while keeping the overhead for local partners low.
Dolly Parton came up with the initial local structure for her region in Tennessee and David Dotson of Dollywood scaled the operation using this decentralized model. Very impressive structure that makes this sustainable.
Here’s the interview where he talks about it: https://www.endbookdeserts.com/dolly-parton-imagination-libr...
PM2.5: 2.6 µg/m³
PM10: 0.5 µg/m³
VOC index: ~97
CO₂: ~505 ppm
That's the baseline. When we're cooking, like now, it's: PM2.5: 10.8 µg/m³ (AQI ~46)
PM10: 17.5 µg/m³
VOC index: 122
CO₂: 572 ppm
None of those seem particularly concerning. We have a gas range, and 15 min past the end of cooking everything is fine. I must imagine that most people don't live right next to the freeway so they probably have even better air quality except during wildfire season.This is why the massive tok/s you see on many small MoE never manifest as worthwhile outcomes. Sonnet 5.5 can be faster on tasks than DeepSeek 4 Flash Vision simply because it gets there faster.
The slow warming of the water is a good approach. I’ve been using abliteration.ai’s cybersecurity models to pentest my own software and it produces much more usable defensive results. Opus 5.5 can fix everything but the guardrails trigger too easily.
The Fable ban affair was a nightmare for defensive cybersecurity. As far as I’m concerned we’re already in the world of Void Star and Cloudbreaker is here.
My own parents are still relatively young and my in-laws are sceptical enough of other people that it’s currently safe but it won’t be so forever.
Edit: hahaha you guys are funny. “The php hack” is what I meant!
For my part, there is a huge price I’m willing to pay to be able to drink water out of my tap. I didn’t have this always and I really value it.
Anyone motivated to do this for a smaller group would rather just run a private tracker. The overheads are enormous when you’re not negotiating in bulk.
Obviously more interesting if I'd thought to screenshot prior to going to bed but you can have multiple tabs in a 'stream' and they can be terminal or web tabs. The relay (which I'm pentesting in the screenshot) transports the AI chats to a paired phone where they're adapted to have a common UI across the Claude / Codex / Oh My Pi that I use. You can obviously see the Arc lineage to the ideas here.
Some 15 years ago people would say "Why would crypto miners even sell you hardware? Why not just keep it for themselves and mine the crypto?" and then it turned out that after you paid for a device you'd often get it slightly used with the cryptocurrency going to the hardware miner. The models are unbelievably good.