I think at a certain scale we're talking about switching to local trained models which don't have the same operating costs as running a frontier model for OCR. That would reduce the ongoing costs significantly. Might take longer than 30 seconds to read each receipt if you run multiple passes to ensure accuracy, but could run 24/7/365 without the same tax and administration overhead of humans.
Spherical cows aside though, I do agree with you that I should not consider scalability as a given.
I suppose if we had access to a public data set like this receipt bank, programmers could time themselves setting up a solution with off the shelf OCR algos. If they could clock in at under 10 hours they could advertise themselves as being "just as good as an LLM, but significantly cheaper." Downside for the managerial class that wants generative algos for the complete lack of legal protections.
>>So I told Codex “we have unlimited tokens, let’s use them all,” and we pivoted to sending every receipt through Codex for structured extraction. From that one sentence, Codex came back with a parallel worker architecture - sharding, health management, checkpointing, retry logic. The whole thing. When I ran out of tokens on Codex mid-run, it auto-switched to Claude and kept going. I didn’t ask it to do that. I didn’t know it had happened until I read the logs.
----
For anybody still thinking my goodness, how wasteful is this SINGLE EXAMPLE: remember that all of the receipts from the article have helped better-train whichever GPT is deciphering all this thermalprinting.
For a small business owner (like my former self), paying $1500 to have an AI decipher all my receipts is still a heck of a lot cheaper than my accountant's rate. It would also motivate me to actually keep receipts (instead of throw-away/guessing), simply to undaunt the monumental task of recordskeeping.
----
>>But the runs kept crashing. Long CLI jobs died when sessions timed out. The script committed results at end-of-run, so early deaths lost everything. I watched it happen three times. On the fourth attempt I said “I would have expected we start a new process per batch.” That was the fix ... Codex patched it, launched it in a tmux session, and the ETA dropped from 12 hours to 3. Not a hard fix. Just the kind of thing you know after you’ve watched enough overnight jobs die at 3 AM.
>>11,345 receipts processed. The thing that was supposed to take all night finished before I went to bed.
Your name is well earned! "can't cure cancers" is impressively counterfactual [0] as 5 year survival of cancer diagnosis is up over almost all categories. Despite every cancer being a unique species trying to kill you, we're getting better and better at dealing with them.
In my experience, most people with cancer that I know simply oscillate between having life-threatening active cancer/tumors and remission.
I don't know any case where people have detectable cancer and it's just being managed, I think that's more the exception than the rule.
For my girlfriend, when she was in her last stages they had to do that (try to slow down/manage the cancer instead of remove it), but that was already palliative care and she died soon after. Also, the only reason they didn't try removing the tumor is because the specific location in the brain (pons) is inoperable.
Yes, we're getting better at treating cancers, but still if a person gets cancer, chances are good the thing they'll die of is cancer. Middling results.
Because we're not good at curing cancers, we're just good at making people survive better for longer until the cancer gets them. 5 year survival is a lousy metric but it's the best we can manage and measure.
I'm perfectly happy investing roughly 98% of my savings into the thing that has a solid shot at curing cancers, autoimmune and neurodegenerative diseases. I don't understand why all billionaires aren't doing this.
If we knew that we probably wouldn’t need AI to tell us.
But realistically: perhaps by noticing patterns we’ve failed to notice and by generating likely molecules or pathways to treatment that we hadn’t explored.
We don’t really know what causes most diseases anyway. Why does the Shingles vaccine seem to defend against dementia? Why does picking your nose a lot seem to increase risk of Alzheimer’s?
That’s the point of building something smarter than us: it can get to places we can’t get on our own, at least much faster than we could without it.
I don’t think that lack of intelligence is the bottleneck. It might be in some places, but categorically, across the board, our bottlenecks are much more pragmatic and mundane.
Consider another devastating disease: tuberculosis. It’s largely eradicated in the 1st world but is still a major cause of death basically everywhere else. We know how to treat it, lack of knowledge isn’t the bottleneck. I’d say effectively we do not have a cure for TB because we have not made that cure accessible to enough humans.
Flying is a bad example because airlines are a thing and make flying relatively accessible.
I get your point, but I don’t think it really matters. If a cure for most (or all) cancers is known but it’s not accessible to most people then it is effectively nonexistent. E.g it will be like TB.
> We have treatments (cures) for TB
TB is still one of the top 10 causes of death globally.
Things like antibiotics are plenty accessible - 3rd world countries are literally overusing and misusing antibiotics to the point of causing drug resistance in TB. "Effectively we do not have [thing] because we have not made that [thing] accessible to enough humans" is an exercise in goal-post moving.
About 15% of people over the age of 15 are illiterate, but it'd be silly to say "effectively we don't have literacy", even in a global context. Depending on the stat, 1 in 10 don't have access to electricity, but electricity has been in 50% of American homes for over 100 years.
The reality is that the future is unevenly distributed. AI and more broadly technology as a whole, will only exacerbate that uneven distribution. That's just the reality of progress: we didn't stall electrifying homes in NYC because they didn't get electricity in Papua New Guinea.
If AI discovers a cure for cancer, it may be incredibly unevenly distributed. Imagine it's some amp'd-up form of CAR-T, requiring huge resources and expenses, but offering an actual cure for that individual. It'd be absurd to say we couldn't consider cancer cured just because the approach doesn't scale to a $1 pill.
> As an example, in the UK in 2013 the cost of standard TB treatment was estimated at £5,000 while the cost of treating MDR-TB was estimated to be more than 10 times greater, ranging from £50,000 to £70,000 per case.
I pulled this from Wikipedia. It does not look like TB treatment is “plenty affordable”.
If the issue is with the semantics of the word “cure” that’s not a hill I’ll die on, but can you see how knowing how to cure something and actually curing something are two vastly different things?
If you told someone a cure for cancer existed but there’s literally no way they could afford it, that sounds a lot like the cure effectively doesn’t exist for that person.
So I’ll posit that the weirdness of such a statement depends entirely on your audience.
If you’re one of the people likely to be able to afford such a cure, it might sound nonsensical.
I’ll also note that I intentionally selected a term with a more narrow definition “effective existence” vs a more general term “existence”. E.g. something can be true in general but effectively false in practice.
It gives me pause that most people drive cars or are willing to sit in one for more than 20 minutes a week.
But people accept the status quo and are afraid to take a moment’s look into the face of their own impending injury, senescence and death: that’s how our brains are wired to survive and it used to make sense evolutionarily until about 5 minutes ago.
Ah, yes: "well, we can't cure cancer or autoimmune and neurodegenerative diseases, but I'm willing to invest basically all my money into a thing that's...trained on the things we know how to do already, and isn't actually very good at doing any of them."
...Meanwhile, we are developing techniques to yes, cure some kinds of cancer, as in every time they check back it's completely gone, without harming healthy tissue.
We are developing "anti-vaccines" for autoimmune diseases, that can teach our bodies to stop attacking themselves.
We are learning where some of the origins of the neurodegenerative diseases are, in ways that makes treating them much more feasible.
So you're 100% wrong about the things we can't do, and your confidence in what "AI" can do is ludicrously unfounded.
Every doctor and researcher in the world is trained on things we already know how to do already.
I’m not claiming we haven’t made a dent. I’m claiming I’m in roughly as much danger from these things right now as any human ever has been: middling results.
If we can speed up the cures by even 1%, that’s cumulatively billions of hours of human life saved by the time we’re done.
But what they can do, that AI can't, is try new things in measured, effective, and ethical ways.
And that hypothetical "billions of hours of human life saved" has to be measured against the actual damage being done right now.
Real damage to economy, environment, politics, social cohesion, and people's lives now
vs
Maybe, someday, we improve the speed of finding cures for diseases? In an unknown way, at an unknown time, for an unknown cost, and by an unknown amount.
Who knows, maybe they'll give everyone a pony while they're at it! It seems just as likely as what you're proposing.
this[0] page makes it seem 500~1000 cycles till 80% starting performance is common. So if you were charging it every other day from a 40~50 mile round trip commute, after 3~5 years you'd go to charging it every day.
As described there, this assumes slow overnight charging, and latest generation of batteries (not sure how viable that was the time of EV1).
Even LiOn batteries have charging patterns as the blocker to adoption, which means that practically, you'd get cars with less than 50% capacity by 2 years.
Also, not like it just keels over and dies, that's just the 80% performance criteria. Most people wouldn't need to replace the batteries at that point.
Do you think it's odd you only listed companies with already existing revenue streams and not companies that started with and only have generative algos as their product?
/Owning/ the truck would be imaginary/aspirational as you imagine yourself using it often enough to justify the expenditure. If it's for trips that are 3~4 times a year it probably makes more sense to rent.
Fortunately other chemistries are making their way down the learning curve lithium's been on, and after a few moment reflection it would be astounding if the same types of storage that are optimal for mobile consumers also happen to be optimal for grid balancing, as daily grid balancing can be done pretty reasonably with a c rating[0] of about .2 or even .1, and as long they aren't so heavy they need to be placed on massive foundation piles, weight is not really a factor. Just need to make it as cheap as possible.
No idea who that is, but it's not particularly challenging to realize evolution doesn't overall favor empathy -even if it played some role-, sometimes it's the full opposite, sometimes is punished ("no good deed goes unpunished"), firemen are the most prone to burns, the equivalent it's true for many other altruistic endeavors, including rare occasional ones unlike firemen.
It's also not particularly challenging to see society lacks any intrinsic defence from the most ruthless and greedy from advancing in any given power structure inside of it, it's a long term damage so it's abstracted away while more immediate issues take presedence, it's in our DNA to give too priority to immediate threats, while long term problems such as this don't make the top 10 (another example being climate change, etc)
I'm juggling a baby approaching lunch time at the moment, so I can't go into too much depth on this paper[0] I found 40 seconds ago, but the conclusion seems to think that solar panel EROIE in siwtzerland is somewhere between 7 and 10, which as a proxy for carbon intensity, /probably/ means it will actually result in net carbon reduction.
It has some elements of standard tragedy of the commons, suppose nearly everybody is following the same rotation scheme, and holding some antibiotics out of rotation entirely as emergency last resort type of deal. Anybody who isn't playing along can be quite confident they're the only person using the emergency drugs and that they'll work quite well.
You got me wondering. Supposing the average post is 10 words, and a typical page of text is 250 words, that would only be ~50 pages of text a day over the last 10 years. Which I don't think I manage, but over 20 years I am probably in that window.
reply