This is misleading. They had 4.5 which was a new scaled up training run. It was a huge model and only served to pro users, but the biggest models are always used as teacher models for smaller models. Thats how you do distillation. It would be stupid to not use the biggest model you have in distillation and a waste since they have the weights.
The would have taken some time to calculate the efficiency gains of pretraining vs RL. Resumed the GPT-4.5 for whatever budget made sense and then spent the rest on RL.
Sure they chose to not serve the large base models anymore for cost reasons.
But I’d guess Google is doing the same. Gemini 2.5 samples very fast and seems way to small to be their base pre train. The efficiency gains in pertaining scale with model scale so it makes sense to train the largest model possible. But then the models end up super sparse and oversized and make little sense to serve in inference without distillation.
In RL the efficiency is very different because you have to inference sample the model to draw online samples. So small models start to make more sense to scale.
Big model => distill => RL
Makes the most theoretical sense for training now days for efficient spending.
So they already did train a big model 4.5. Not using it would have been absurd and they have a known recipe they could return scaling on if the returns were justified.
Even worse for non-American stockholders of American companies - the IRS charges a 30% foreign withholding tax on dividends. If you ban stock buybacks in favour of dividends, it’s a big tax increase on foreigners, so US stocks lose a whole pile of value for American stockholders when foreigners dump American equities until the ROI equalizes. (Roughly 20% of US equities are foreign-owned.)
It's a lot lighter: a stack trace takes a lot of overhead to generate; a result has no overhead for a failure. The overhead (panic) only comes once the failure can't be handled. (Most books on Java/C# don't explain that throwing exceptions has high performance overhead.)
Exceptions force a panic on all errors, which is why they're supposed to be used in "exceptional" situations. To avoid exceptions when an error is expected, (eof, broken socket, file not found,) you either have to use an unnatural return type or accept the performance penalty of the panic that happens when you "throw."
In Rust, the stack trace happens at panic (unwrap), which is when the error isn't handled. IE, it's not when the file isn't found, it's when the error isn't handled.
To defend Wilkins, it was John Randall, the director of the lab Wilkins and Franklin both worked in, who probably intentionally pitted them against each other to mess with or motivate Wilkins. Wilkins was possibly the most honorable out of all five people involved in the situation.
Wilkins was "second-in-command" to Randall, developed the DNA structure project, and convinced Randall to assign more people to work on it. Randall then hired Franklin, reassigned Gosling, the graduate student who had been working with Wilkins, to Franklin, and told Franklin that Wilkins would simply be handing over his data to her and that she would subsequently have full ownership of the project. Randall didn't tell Wilkins any of this of course, so a lot of hard feelings developed between Franklin and him. The situation got worse when Wilkins tried to get sample from external collaborators to continue working on the project himself and Randall forced him to hand over one of the samples to Franklin. Franklin finally got sick of Randall herself and left, leaving Randall to turn over all the data to Wilkins, who then went to talk about his pet research interest with Crick, a personal friend of his. Wilkins then recused himself from Crick's paper, feeling he hadn't contributed enough to it. He also worried publicly to others that maybe he had been unkind and driven Franklin out, having minimal insight into Randall's tactics, which are unfortunately common in the field. When they're being used on you by someone skilled in them, it's often hard to realize, and you end up being resentful of the person you're being pitted against until one of you leaves and you suddenly have clarity because the stress of the situation is suddenly reduced.
> But here’s the most telling data point from the school’s National Student Clearinghouse data: Over 60% of WKU-admitted applicants who don’t attend WKY don’t enroll anywhere. Not at a competitor, not at a community college — nowhere. Colleges are still competing with one another, but increasingly, they’re competing with the labor market itself.
Labor shortages are leading to college credentials not being needed. This is objectively good (as college debt and the time opportunity cost is avoided for an unnecessary credential), and hopefully will continue as demographic dynamics continue in the US. Is it good or bad the US has many colleges that will close because they are no longer needed due to a slowly declining fertility rate? It just is.
Anyone who has lived through a market correction (the tariff announcements in early April this year being a recent example, though there have been far worse) should be able to see that market prices do not always accurately reflect even the consensus view of value (which itself can be wrong). As people are forced to de-lever, everything goes down at once, often by very similar amounts, even though it cannot be possible that everything suddenly lost the same amount of value simultaneously.
To quote Richard Bookstaber, "The principal reason for intraday price movement is the demand for liquidity... the role of the market is to provide immediacy for liquidity demanders. ...market crises... are the times when liquidity and immediacy matter most. ...the defining characteristic is that time is more important than price. ...diversification strategies fail. Assets that are uncorrelated suddenly become highly correlated, and all positions go down together. The reason for the lack of diversification is that in a high-energy market, all assets in fact are the same.... What matters is who holds the assets." (from A Framework for Understanding Market Crises, 1999)
Was the market drop an accurate reflection of the value that would have been destroyed by those tariffs, discounted by the probability that they would have been enacted as drafted? Nobody knew then, and I maintain that nobody even knows now. That was not the calculation that was being made.
"Federal funding typically covers 80% of bus purchases, with agencies responsible for the remainder."
Well, there is your answer. The one making the purchase isn't the one primarily paying for the purchase. This makes them less sensitive to pricing.
Kinda like how expensive healthcare is since it is paid for by insurance.
Or how you don't care how much you put on your plate or what you choose to eat at an all you can eat buffet.
The second you detach the consumer from the price of something, even through an intermediary such as health insurance, that is when they stop caring about how much something costs, and so the price jumps.
Practically stars are mostly burned out in another 50 billion years and radioisotopes that produce a heat gradient will also be mostly decayed by then. Eventually good tidal energy situations like
will end as well since this kind of situation changes the orbits. So energy for life and usable thermal gradients will disappear even if entropy will continue to increase for a long time -- for instance, black holes will be slowly inspiralling and crashing into each other resulting in huge entropy increases on paper.
This article reminds me of my early days at Microsoft. I spent 8 years in the Developer Division (DevDiv).
Microsoft had three personas for software engineers that were eventually retired for a much more complex persona framework called people in context (the irony in relation to this article isn’t lost on me).
But those original personas still stick with me and have been incredibly valuable in my career to understand and work effectively with other engineers.
Mort - the pragmatic engineer who cares most about the business outcome. If a “pile of if statements” gets the job done quickly and meets the requirements - Mort became a pejorative term at Microsoft unfortunately. VB developers were often Morts, Access developers were often Morts.
Elvis - the rockstar engineer who cares most about doing something new and exciting. Being the first to use the latest framework or technology. Getting visibility and accolades for innovation. The code might be a little unstable - but move fast and break things right? Elvis also cares a lot about the perceived brilliance of their code - 4 layers of abstraction? That must take a genius to understand and Elvis understands it because they wrote it, now everyone will know they are a genius.
For many engineers at Microsoft (especially early in career) the assumption was (and still is largely) that Elvis gets promoted because Elvis gets visibility and is always innovating.
Einstein - the engineer who cares about the algorithm. Einstein wants to write the most performant, the most elegant, the most technically correct code possible. Einstein cares more if they are writing “pythonic” code than if the output actually solves the business problem. Einstein will refactor 200 lines of code to add a single new conditional to keep the codebase consistent. Einsteins love love love functional languages.
None of these personas represent a real engineer - every engineer is a mix, and a human with complex motivations and perspectives - but I can usually pin one of these 3 as the primary within a few days of PRs and a single design review.
The IPO "pop" is not captured by banks: it's captured by the banks customers that pre-buy at the IPO price.
Basically, before an IPO, the underwriters take the company on a "roadshow" in which they pitch the IPO to potential buyers.
There's a hierarchy of these: the best are very large buyers that place large orders and trade seldom. Pensions, sovereign wealth funds, etc.
Those buyers then make offers ("I'll buy 50MM at $100"), which the bank uses to set the IPO price. The bank then gives them an allocation.
If you're a high (10MM+) net worth individual that banks with one of the underwriters, you can often get an allocation in an IPO. The richer you are, the more of an allocation you can get.
When an IPO pops, it's these people that get the benefit.
The benefit for the company is that the stock is owned by prime people the bank selected: you crucially _don't_ want to just sell to the highest bidder if they are going to dump the stock immediately after the pop (or that's the theory, at least). They have stable shareholders with a vision aligned with management.
The benefit to the bank is that they get to reward their customers with access to profitable trades--but the bank itself does not profit.
This will be one of the big fights of the next couple years. On what terms can an Agent morally and legally claim to be a user?
As a user I want the agent to be my full proxy. As a website operator I don’t want a mob of bots draining my resources.
Perhaps a good analogy is Mint and the bank account scraping they had to do in the 2010s, because no bank offered APIs with scoped permissions. Lots of customers complained, and after Plaid made it big business, eventually they relented and built the scalable solution.
The technical solution here is probably some combination of offering MCP endpoints for your actions, and some direct blob store access for static content. (Maybe even figuring out how to bill content loading to the consumer so agents foot the bill.)
"electrons flow through the respiratory chains of the respective feedback controllers like sand in the hourglass that determines when balance must be restored"
Wow, that is my new favorite sentence from any paper ever, replacing Mark Thomas' equally epic:
"What it begins to suggest is that we’re looking at a Lord of the Rings-type world" from the legendary meeting at the Royal Society in London 2012/13.
The federal government alone spends $1.9 trillion annually on healthcare. That's enough to buy almost a million Tomahawk missiles every year. The total production will be around 9,000 missiles over 46 years, or less than 200 per year. We do not meaningfully choose between paying for healthcare domestically and blowing up foreigners. Even overthrowing Iraq's government and trying to make it a democracy only cost about $2.4 trillion over 10 years.
In high school science class we boiled water using standard printer paper: fold a sheet into a box, put it on a stand over a bunsen burner, fill it with water, turn up the bunsen burner.
Not only does it not burn but it retains more of its structural integrity.
I'm pretty sure the person who wrote this has never run pricing research for a brand. Short answer, they can ignore Gabor-Granger because their cost base is so low compared to their revenue, so they'd be looking at Van Westendorp's Price Sensitivity Meter to set a benchmark for where the pricing probably lands, and a conjoint study to understand the value of different elements for segmenting different versions of the product at different price levels.
Obviously positioning, who they're positioning against, how they communicate that, the level to which they're known amongst the market etc all feed in to this, but that'd be a decent starter for ten.
This is an overly simplistic version of where to go with pricing for a brand like this, but that's where I'd begin with creating pricing for them.
My father was on chemotherapy with fludarabine, a dna base analog. The way it functions is that it is used in DNA replication, but then doesn’t work, and the daughter cells die.
Typically, patients who get this drug experience a lot of adverse effects, including a highly suppressed immune system and risk of serious infections.
I researched whether there was a circadian rhythm in replication of either the cancer cells or the immune cells: lymphocyte and other progenitors, and found papers indicating that the cancer cells replicated continuously, but the progenitor cells replicated primarily during the day.
Based on this, we arranged for him to get the chemotherapy infusion in the evening, which took some doing, and the result was that his immune system was not suppressed in the subsequent rounds of chemo given using that schedule.
His doctor was very impressed, but said that since there was no clinical study, and it was inconvenient to do this, they would not be changing their protocol for other patients.
A lot of people are saying that a) AI generates slop that no one needs, and b) AI is putting human artists out of work.
If the machine can do art that's indistinguishable from human art, and art is the soul of humanity, then the machine may have a soul? I've told the machine to create art, I've showed the art to humans, and the humans were touched by it. It evoked an emotion, like art is supposed to.
My personal anecdote: I've used a diffusion model to generate a short video based on a 50 year old photograph, the only photo my dear friend has of his late father that he never got to know. The 10-second video showed the man lifelike, happy and smiling, generated from a photo on which he looked morose. My friend was brought to tears when I showed it to him.
Scott Adams' revolution was to get users to give him plot lines.
He was the first to publish an open way to communicate with him in order to out the corporate crazies, and readers did in droves, explaining the inanity of their workplace and getting secret retribution for stuff they clearly couldn't complain about publicly.
A good percentage of youtubers and substackers today actively cultivate their readership as a source of new material. They're more of a refining prism or filter for an otherwise unstated concerns than a source of wisdom.
Doing this seems to require identifying with your readers and their concerns. That could be disturbing to the author if the tide turns, or to the readers if they find out their role model was gaming them or otherwise unreal, but I imagine it is pretty heady stuff.
I hope he (and anyone facing cancer) has people with whom he can share honestly, and has access to the best health care available.
I can't just treat `bar` the same as `foo` because it doesn't give me a String, it might have failed to give me a String. So I need to give it special handling to get a String.
async fn qux() -> String
This also doesn't give me a String. It gives me a thing that can give me a String (an `impl Future<Output=String>`, to be more specific), and I need to give it special handling to get a String.
All of these function have different colours, and I don't really see why it's suddenly a big issue for `qux` when it wasn't for `bar`.
That depends. Look it up. You will find there is a point where it switches. Normally the body (of both baby and mother) will protect the mother. Something goes wrong or just gets too far "out of spec"? Miscarriage. After a few months, the body goes so far as to sedate the mother and child before terminating the pregnancy. There is research claiming it actually shuts down the baby's nervous system before decoupling.
But about a month before birth things switch around. The womb partially disconnects from control systems of the mother's body and ... there's an extremely scary way of pointing this out I once heard from a medical professor: "you know just about the only thing a human body can still do when it's decapitated? It can give birth"
In less extreme circumstances, you actually have a switch in your circulatory system ... when pregnancy gets to this point and the mother's body loses power, it will initiate a rapid birthing process, and start shutting down organ after organ to give birth with the remaining power. That includes, eventually, the brain. Only the heart, lungs, liver and womb will remain operational. The body will shut down blood flow to the brain to continue giving birth. Once shut down it cannot be turned back on. So this kills the mother, despite the body remaining functional, in some reported cases, for over an hour, and is something gynaecologists get trained to prevent from happening.
Given how common it was even a century ago for women to die giving birth, one wonders how often this mechanism was involved.
This is interesting to me at the margins, because one of the things I learned when my wife got pregnant the first time was that the womb is not exactly the warm cradle of nurturing that I had always (without thinking much about it) imagined, but in many ways a blast door or containment vessel to protect the mother (host) from the fetus (roughly, xenomorph) that would otherwise explode like an aggressive parasite (killing them both).
So I mean, you probably don't want to have any leaks or weak stitches in your uterus transplant...
The anti-DEI argument is that modern racial disparities are predominantly caused by economic circumstances, e.g. black people are more likely to be poor and then less likely to have to startup capital to start their own business or be able to afford to attend a high status university. The same applies to white people who don't have affluent parents. "White people who grew up poor" are under-represented at the top of society.
So the underlying problem here is economic opportunity, not race. To fix it you need to e.g. make it easier for someone without rich parents to start a business by lowering barriers to entry and regulatory overhead on small entities. That allows both poor black people and poor white people to get ahead without discriminating against anyone, but still reduces the racial disparity because black people are disproportionately poor.
It's basically Goodhart's law. Because of the existing correlation between race and poverty, continuing racial disparities are a strong proxy for insufficient upward mobility, but you want to solve the actual problem and not just fudge the metric through race quotas etc.
The would have taken some time to calculate the efficiency gains of pretraining vs RL. Resumed the GPT-4.5 for whatever budget made sense and then spent the rest on RL.
Sure they chose to not serve the large base models anymore for cost reasons.
But I’d guess Google is doing the same. Gemini 2.5 samples very fast and seems way to small to be their base pre train. The efficiency gains in pertaining scale with model scale so it makes sense to train the largest model possible. But then the models end up super sparse and oversized and make little sense to serve in inference without distillation.
In RL the efficiency is very different because you have to inference sample the model to draw online samples. So small models start to make more sense to scale.
Big model => distill => RL
Makes the most theoretical sense for training now days for efficient spending.
So they already did train a big model 4.5. Not using it would have been absurd and they have a known recipe they could return scaling on if the returns were justified.