> but they will probably soon have a GPT-bio that they license to Pfizer and a GPT-quant that they license to Citadel, rather than keeping those valuable capabilities in the public models
Nein. Neither Pfizer nor Citadel are going to use a hosted model - their whole business is IP. They will invest in self hosting models. Self hosting is becoming a big deal in the industries you're talking about.
Why?
Neither Pfizer nor Citadel wants to pay a supplier/vendor to build a core capability that the supplier/vendor will turn around and sell to their competition.
Sure, and OpenAI will be happy to let them self-host for a large fee. However I doubt that Pfizer will be developing their own frontier models for biological research.
> If you ask an LLM to give you a 500-word summary of quantum physics, it'll give you an oversimplification that probably leans on a hodgepodge of pop-sci metaphors
What would a double PhD in quantum physics provide differently if you asked them for a 500-word summary of such a complex field? What would the human do very differently? I have an Associate Professor from CalTech who teaches quantum physics there and I will have them review your suggestions, so don't hesitate out of concern.
> I'd just really like to see at least one of these to be accompanied by a statement saying what are the kinds of problems the author can now confidently solve that they couldn't before.
I don't get why there's this overwhelming dislike for LLMs on HN. Everytime I make a comment on how insanely productive it has made me, I get downvoted to hell that leads me to be throttled by HN for hours at which point I can't engage in the discussion anymore. I say this in advance because if you post a comment and don't see a response from me until the next day, that's what's going on.
So here's one of a dozen ways LLMs have helped me to learn, execute and deploy ideas rapidly.
MiniPCs that used to cost no more than a burrito, Raspberry Pis, and useful hardware like that are extremely expensive right now: so I have to be creative in finding replacements. I have been able to use GPT5.5/GPT5.6 Sol/Gemini 3.5 Pro/Opus 4.8 to locate cheap ($5) routers and very cheap repeaters that can be reprogrammed to run Linux on them.
This is an extremely intensive, laborious process that requires:
1. Flashing the device over Qualcomm EBL.
2. Validating that I didn't corrupt the 4GB eMMC, then partitioning it.
3. Doing multiple gated commands that verify step by step that the previous command worked correctly and had the desired effect.
4. If not, take remediation steps, failing which alert me so we can do a spike.
5. Remapping certain hardware and Flashing over a working Debian.
6. Install DropSSH + keys and validating that.
7. Installing the scripts, etc.
Since these are not meant to be used this way, and I'm repurposing them, ordering these cheap Wi-Fi routers and flashing them is not a repeatable process. Each one is slightly different from each other. You really can't script it, not reasonably.
OpenCode using the above models has cost me $5 worth of tokens so far to reflash 10 of these repeaters, at $5 each, into Microservers that do my bidding. They have 1 GB of RAM, 4 GB of eMMC, can do USB OTG and being routers have WiFi and BT. A compatible hardware today would cost me ATLEAST $50+ each.
Thanks to the LLMs for walking me through discovering this detail, holding my hands through the process and handing me over these working microservers.
Wait, I can purchase $2-$4 "find hub" compatible beacons all day and have been for months. I have even made my own using a CR2032 + ESP32-C3 and they work fantastic with "find hub" - is this just an official version of that or something else entirely?
The more I read about articles like these, I grow even more convinced that the education industry as a whole, regardless of the country or stature, has kind of lost their plot.
Sometimes I even wonder if this is the outcome of sheer laziness, fear or both.
When I speak to professors, teachers and students - It's also surprising that "top tier" institutions are fighting AI harder (outside extremely specific courses like Harvard’s flagship CS50, MBA courses at Wharton (UPenn) and MIT) while "bottom tier" institutions are completely embracing it and rebuilding their curriculum around it. One CS professor at a "bottom tier" CSU mentioned to me that he's actively going completely "open book using AI" - students are allowed to use anything they want from Claude to Codex to OpenCode to finish assignments but the assignments have now changed from "blurt out quicksort" to "let's sort N natural numbers in a cache efficient way using least amount of resources". I would hire the latter over the former anytime. I am tired of interviewing candidates who can shit out quicksort before I can even finish my sentence but stare at me dumbfounded when I ask them to sort people's name serialized in unicode.
In my opinion, the bulk of traditional education has been a mix of memorization of facts, knowing inference rules, and applying inference rules to those facts, coupled with recall.
Before the age of LLMs, only very expensive-to-build rule-based expert systems were able to replicate that functionality. Humans were just simply cheaper and way more reliable.
In 2026, Frontier models are exceedingly, across the board, across industries, breaking records and challenging those notions on cost, capability and sophistication.
Trying to replicate how education used to operate pre-frontier LLMs is like forcing people to farm by hand in the age of automated tractors that have LIDAR, Vision, RTK on board.
I'm not discounting other elements of learning like collaborative debate, clinical/lab work, Socratic reasoning, emotional intelligence and the development of a professional network - but I argue these skills are not limited to a school or university setting. Infact, a lot of this is distorted in a school or university setting compared to the real world.
I have been filing my own taxes, including complexities like equity, real estate and business income for over a decade now, so it's not just a simple 1040 and 540. Reading through IRS documentation, talking to EAs and CPAs to fill in ambiguities and gaps was pretty expensive in terms of time and money.
Both Claude Opus and GPT 5.5 now, as of 2026, answer all my tax questions correctly. Those thousands of dollars in time and money I had spent has been replaced by a single $20 subscription. Unless tax codes drastically change every year, that $20 is a one time cost.
If I were to begin my tax journey in 2026, I would never had to spend those dollars: dozens of tax professionals are out of a job - all their education is for nothing.
It's entirely irrelevant whether they passed their exams by carving answers out on granite, taking their exams in a jail cell on an island proctored by Catholic Nuns under the watchful guard of automatic machine guns manned by T-1000s, I just don't need them anymore. For my usecase, these humans and their credentials provide 0 additional value over a one-time $20 expense, no matter how complicated it was for them to get their credential and what complex interpretive dance they had to do to impress the people awarding their grades.
All this "reject AI, do it by hand" is just insanity at worst and laziness at best.
That said, I'm personally unsure what the future of education in the age of tractors is like but removing weeds, planting seeds, watering them, all by hand is certainly not it.
It's my understanding that most countries - developed or developing - are bottlenecked on educating their masses due to lack of qualified teachers.
That's a bandwidth problem. Having those teachers listen to oral arguments from a handful of students is not solving the core bandwidth problem. All these shenanigans is doing a disservice to the public.
> In my opinion, the bulk of traditional education has been a mix of memorization of facts, knowing inference rules, and applying inference rules to those facts, coupled with recall.
These are important skills to develop, even in a world with LLMs.
> All this "reject AI, do it by hand" is just insanity at worst and laziness at best.
If a 3rd grader complained about having to learn multiplication because calculators exist, would you agree with them?
Why do we make kids learn how to do math if they can pull out their phone and open the calculator app? Because learning how to learn is important and there is value in understanding how the answer is produced, even if you have a machine that can produce it for you.
Universities can teach a separate class on how to use LLMs. The existing classes should be focused on teaching their existing subject matter, not becoming an extension of a how-to-LLM class
> Why do we make kids learn how to do math if they can pull out their phone and open the calculator app? Because learning how to learn is important and there is value in understanding how the answer is produced, even if you have a machine that can produce it for you.
I actually appreciate this question very much because this question is extremely pertinent to our discussion. Before I continue my response, I want to turn around and ask you:
1. What's your definition of "learn how to do math"?
i. Would it be sufficient if they proved they understood what addition, subtraction, multiplication and division was? Or do they need to be able to correctly calculate what 4592 * 314 is? What if they could show you using diagrams of squares and rectangles *why* (a + b)** 2 expands to the equation it does but refused to chart out results for various values of a and b?
ii. Would you fail a student who could reliably chart out results for (a + b)** 2 given various values of a and b, but failed to explain that using diagrams of squares and rectangles?
iii. Would you fail a student who could reliably add 1 + 3 or 7 - 5 or 7 * 5 but fail to compute 4592 * 314?
iv. Is a student who could reliably add 1 + 3 or 7 - 5 or 7 * 5 but fail to compute 4592 * 314 at the age of 5 superior or inferior to a student who is capable of doing the same at age 7?
2. If an exam is composed of tables of long division and your score on the test boils down to your ability to how many of those long divisions you can complete in 45 minutes, does the person who scores the highest have a superior understanding of division than the slowest student who suffers from mental fatigue due to an underlying, undiagnoised health condition?
3. When you say "understand how the answer is produced", what level of abstraction is acceptable to be not considered as cheating? Do they need to understand how the calculator physically computes the floating-point math, or is pushing the button enough? If pushing the button is cheating, why isn't using a base-10 shortcut algorithm also cheating? What if the button pusher explained to you precisely how IEEE 754 worked, how a digital calculator works and then refused to do long division citing it was a complete waste of their time and yours? Would you refuse to let them pass the class or fail them unless they yielded to your demands and complied with your specific definition of learning?
Is the math class also doubling as ability to pass compliance and behavioral standards or is it purely a test of mathematical ability?
4. If a student uses an LLM to generate the boilerplate code for a script, but can perfectly explain the architecture, debug the logic, and scale the deployment, have they failed to 'learn how to learn' just because they didn't manually type the syntax?
Educational resources and classroom time is a zero sum game. We cannot afford to educate everyone if they all require 1:1 coaching from a qualified human in a room that has limited space. Time in a day is zero sum as well. Every minute someone spends on doing the 135th long division is a minute they're not spending thinking if solving long division problems is a meaningful differentor in their long term success - whether it's the long division that will make them wealthy, happy and successful or something else entirely?
> i. Would it be sufficient if they proved they understood what addition, subtraction, multiplication and division was? Or do they need to be able to correctly calculate what 4592 * 314 is? What if they could show you using diagrams of squares and rectangles why (a + b)* 2 expands to the equation it does but refused to chart out results for various values of a and b?
I’m not an educator by profession but I’ve done a lot of training and mentoring.
There is no replacement for actually doing the work. People who study something but don’t go through the exercises feel they understand topics better than they do. It’s only when they are put in a position where they have to apply it that the cracks in their understanding are revealed.
This would 100% result in students who thought they could explain multiplication but only had a surface level idea. The knowledge would also be fleeting because actually doing the thing makes the knowledge stick more than just observing and understanding the thing.
So yes, you still need to have the students do the thing.
Replace math with writing and it will be more obvious. It’s easy to look at someone’s writing and critique it. It’s much harder to write well without practicing.
> It’s much harder to write well without practicing
Is the point of writing, to communicate information from one person to another group or something else?
Which of these fail to communicate the need:
i. "Need to go poo poo! NOW!"
ii. "Excuse me Sir. I sincerely apologize for the inconvenience, I didn't mean to intrude but could you, please, point me to the direction of the nearest latrines, or the head as some might say, so I can relieve myself. The burrito didn't sit well with me unfortunately and it's causing me much distress! I must stress, time is of the utmost essence!"
> So yes, you still need to have the students do the thing.
I don't disagree with "There is no replacement for actually doing the work".
What I do disagree with is the exact nature of the work done: when you have access to a gas/electric driven auger, trying to dig a hole by hand with a teaspoon is neither smart nor productive, like learning how to do long division.
> The knowledge would also be fleeting because actually doing the thing makes the knowledge stick more than just observing and understanding the thing
Define "stick".
1. I recall, in school, sitting in classes for days where the teacher went over multiplication tables 5x5 through 99x99. It was painful. I don't recall the top off my head what 93x95 is but I know how to calculate it now and I knew how to calculate it then. Hearing and doing 93x95 on a balckboard was not specially valuable nor additionally insightful nor marginally instructive over 13x19 and we could easily have stopped there and done something else with the remaining time.
2. I also recall, sitting in classes where we did long division by hand. You had to show your work and it was incredibly exhausting. I taught myself special tricks just to finish the long division exercises quickly and just be done with it so I could actually do the things that interested me more like reading manuals about vacuum tubes and how triodes worked.
3. Then there was this phase of "unitary method" which was just batshit crazy. They would ask questions like "if it takes 1 person to cut 1 block of wood 10 hours, how long would it take 10 men to cut the same wood?", and you would have to follow this insane mechanical process of decomposing the problem into its constituents and do all this crazy, laborious steps to arrive at the answer. It made absolutely no sense to me then and it still doesn't. It felt like digging a trench with a teaspoon. I had taught myself symbolic manipulation by then and used the "X, Y, Z" method to rapidly solve the problem. I recall annoying the teacher with "if it takes 1 person 10 days to cross a river, how long would it take 10 people to cross the same river?"
So back to my question - define "stick". What's the window of "stickiness"? Who defines that window and why?
Next:
4. Has long division been sticky for me (I don't recall it at all but I do remember I scored very well on those exercises then because the alternatives would leave me with even less time to do the things I actually liked doing like building mechanical robots)?
5. Exactly how many exercises were necessary and sufficient to prove that I understood long division? 1 page of long division? 5 pages? 100?
6. Do you recall how to do long division? I don't. I can bet you that I can look it up and do it correctly but I won't out of principle: There's absolutely 0 value in knowing how to do long division. It was as useful to me back then as it is now. We should have done something more productive with our time.
"But doing long division laid the very foundation of logic and mathematical ability!" - Nonsense. There were students in my class who excelled at long division but couldn't grok calculus no matter how hard they tried. The only thing that being able to do long division proves is that you can pay attention long enough and maintain state. There are far more exciting, interesting ways of proving that ability.
If I need to learn long division, I always had the ability to look it up, do some exercises, verify I understand it correctly and complete the task I was relying on long division for. Before the age of LLMs, one might argue that I needed to have some idea that long division was the tool I needed for solving my problem, and thus, having learned long division was valuable. In 2026, I can tell my desired state to a frontier model and it will propose various ways to solve it, then solve it for me if I so desire and walk me through the solution step-by-step if I want it to.
So no, long division, "unitary method" and all such matters were and are nonsense, a waste of time and we should replace them with something more meaningful. If we can't come up with that, we should atleast let the kids go outside and play games they like.
So again - what's this worry about knowledge being fleeting? Do either you, I or my teacher who taught me long division care whether I posses the knowledge of long division in 2026?
Even though I knew long division in 1997, mastered it, excelled at it back then, do I still know it now?
Does my inability to explain long division at a moment's notice in 2026 imply long division hasn't been "sticky"?
If I never, ever knew long division at all, can I correctly learn it now?
Who defines what knowledge is worth fleeting and which one is not? My history teacher and chemistry teacher certainly disagreed on that matter in 1997 and still do!
The only truce they were willing to sign is that history and chemistry are both equally important but given that I am a software engineer by trade and deploying high quality systems to production is how I support my lifestyle, I provably disagree with both of them.
Depends on what skills you want the students to have.
In e.g., philosophy, allowing students to write their assessments with the help of LLMs changes the students’ depth of understanding, as well as their ability to synthesize, formalize, and drum up their own arguments and objections.
These are all useful skills, even in a post-LLM world.
> allowing students to write their assessments with the help of LLMs changes the students’ depth of understanding, as well as their ability to synthesize, formalize, and drum up their own arguments and objections
indeed but LLMs improve these capabilities. Let me explain:
Most educators behave as if using an LLM robs the student of their ability to think. That has not and never been my experience.
You can just tell an LLM to debate, find out inconsistencies and issues in your writing, and challenge you, and it will very well do that!
LLMs for me, cheaply replace teachers who can barely be accessed during office hours because they are so busy doing research or overwhelmed with grading coursework.
In a typical educational setup, I might have half an hour at best with a teacher who looks kindly upon me, and at worst, five minutes with a teacher who doesn't like being around me. An LLM has no emotion, but it has knowledge, it has inference rules, and it has the ability to push back, debate, and find out inconsistencies, bugs and gaps in my thinking.
Behavior that is exactly what I need consistently, in a good teacher.
A frontier LLM is a good teacher who doesn't play games, has no emotion, infinite patience and always available for $20/mo.
> You can just tell an LLM to debate, find out inconsistencies and issues in your writing, and challenge you, and it will very well do that!
Sure, but the point is that you are supposed to be able to find these inconsistencies and issues in your writing. That is the skill.
Indeed, LLMs can be helpful as an alternative to a teacher, if thought of as a sophist (I have even done research on this [0]). But that is not the approach you are proposing.
If you want students to be able to reason about these things on their own, then LLMs doing the hard work of identifying the gaps in their logic is counter-intuitive.
I honestly don't understand what we're discussing/debating? It feels like you think I'm countering your postition but I'm not.
I'm tired of people echoing a common trope that somehow LLMs replace critical thinking or steal learning opportunities from students - they do the opposite. They provide students more opportunities to sharpen their critical thinking and democratize learning.
In your parent post (https://news.ycombinator.com/item?id=49225068), you asserted that "These are all useful skills, even in a post-LLM world.": I agreed with you, and proposed that you could use an LLM to debate you and furthur hone that skill because you certainly don't have unlimited time with a skilled human teacher but for $20/mo, you could debate a frontier model as long as you're awake, all day long.
Your research seems to be supporting what I just said, does it not?
> Indeed, LLMs can be helpful as an alternative to a teacher, if thought of as a sophist (I have even done research on this [0]). But that is not the approach you are proposing.
What according to you, did I propose instead?
I also need further clarification on the term "sophist", because in the US and the UK, it means a person who is "misusing" logic and attempting to deceive their opponents using trickery instead of an honest debate that centers around the objective truth. Their sole objective is to "win" a debate rather than appreciate debate itself or the merits of the topic itself. I believe, even in ancient history, Plato heavily criticized them.
For example: If we were debating whether child labor is wrong, a sophist who was put in a position "for" the topic would then argue there are good reasons why child labor is good. A "true" debator would concede that they could not support child labor and would not debate for the topic.
Is that what you meant by the term "sophist"? Or does "sophist" mean a "Debate Coach" or "Instructor of Rhetoric" or "teacher who excels in the art of debating"? If you meant "Debate Coach" but it sounds too simplistic, may I propose we use "Forensics Instructor" - because the term "sophist" in modern English certainly doesn't communicate that and its use will lead to confusion.
> then LLMs doing the hard work of identifying the gaps in their logic is counter-intuitive
Your use of the phrase "counter-intuitive" is unclear to me. Do you mean to say that:
1. it's actually a good idea that can certainly provide great outcomes but doesn't occur to a lot of people, or
2. that it's not a good idea at all because it just doesn't work?
There are two kinds of engineers in Finland. Some received a more theoretical education at research universities. Most went to more applied institutions that focus more on practical skills. Regardless of what employers say, they generally prefer the theoretical engineers from research universities. Because the higher status of those universities attracts more talented individuals, because theoretical understanding tends to stay valid longer than practical skills, and because it's easier to learn practical skills at work.
The kids who start their studies today are supposed to graduate in 2030, and they will probably retire around 2080. Focusing too much on the skills their early employers might want in the 2030s would be a huge misallocation of resources.
> because theoretical understanding tends to stay valid longer than practical skills, and because it's easier to learn practical skills at work
depends. If you're building a datacenter and need welders to weld two very specific set of metals together, allocating a PhD in metallurgy to do that welding, I argue would be a huge misallocation of resources.
Now your argument seems to be whether the PhD in metallurgy can pivot to doing something else entirely compared to someone who has been narrowly trained to only weld two very specific set of metals together.
If I interpret your take on the matter, your point seems to be that the former is "intellectually superior and more flexible" than the latter.
I dont take that position - I find most humans to be extremely capable. Just because one human was able to grasp the intricacies of metallurgy doesn't necessarily imply they have more capability than someone who didn't.
The fact that employers in Finland can be so very picky says more about the productivity of Finland than the employee's capabilities. At peak productivity, Finland would just not be hiring anyone who could breathe but import labor because there's not enough people in Finland.
First, employers favor applicants from prestigious schools over those from less prestigious schools, because prestigious schools are more selective. That seems to be at least as true in the US as in Finland.
Second, theoretical education is more useful in the long term, because it's a better foundation for learning new skills. It can be a disadvantage when you are trying to get your first job, as you have fewer practical skills potential employers would need right now. Which is why, from an individual perspective, theoretical education only makes sense in higher-tier institutions, where the prestige can offset the initial disadvantage.
I heard your points - you communicated them very well. I failed to communicate that the first, while true, isn't going to matter in a few years time and the second is not true.
I will repeat what I said - if an employer arbitrarily restricts their talent pool to legacy credentials, they are optimizing for signaling rather than raw productivity.
By 2026, there is overwhelming anecdote and evidence that credentials and where you went to school have very little to do with demonstrated performance. (See citations 1-3)
Hence, it's mostly theatrics.
It seems like you're arguing that it's safer to attend prestigious schools because that increases your chances of being employed. I argue that in the age of LLMs, where the cost of learning is very low and speed of building prototypes (and hence learning and improvement) is rapid, it's easier than ever to start your own business.
> First, employers favor applicants from prestigious schools over those from less prestigious schools, because prestigious schools are more selective. That seems to be at least as true in the US as in Finland
This represents a trailing-indicator mindset typical of "MNC".
Yes, legacy corporations still filter by prestige because they want to signal to others they hire only the best, but that's because in extremely large orgs like theirs, it's impossible for a single person's contributions (outside the C-Suite) to have an outsized effect on that org, so they resort to proxy metrics, like credentails, for signalling.
In real life though, credentials have an extremely weak correlation with impact (see citations).
What this translates for an end user, is that the end user doesn't care about the credentials of the person who built the product. I explain that below.
On the frontiers of excellence - startup founders will tell you that hiring people from prestigious schools is a liability over those who might only have a high school degree because the prestigious school graduate is entitled and is looking to build a resume so they can jump ship in 2 years instead of the barely high school degree student who's hungry to prove they are worth investing in.
The selectivity of prestigious schools is gamed heavily in the U.S where you have literally floors of "prep schools" full of trainers telling you which piano lessons to take, which tennis tournaments to play at, which charities to volunteer for and which SAT question types to learn well. Getting into Harvard is less about ability and more about checking off boxes on a checklist.
Notice that absolutely none of this actually has to do with what the student actually wants to do in life and everything to do with impressing the gatekeepers to admissions.
However, make no mistake, the end customer rarely cares about the credentials.
When you watch Netflix or play the XBox, I bet you don't look up the credentials of the production team or the engineers who built it - you enjoy the content and vote with your dollars - and that's real life.
HN is full of people who don't have anything past a high-school degree but they were obsessed with solving one problem well and now they are multi-millionaires and billionaires because they ended up having an outsized impact. There are far more in that bucket who are comfortable but not multi-millionaires (so you never hear about them).
The money is not the end result but proof of impact and this proves my point.
Let's even discard that, citing survivorship bias. Let's reduce my argument to one fact: engineering/code doesn't care about the author's pedigree - all that matters is whether the system compiles, works correctly and scales under load and a prestigious university certainly doesn't have exclusive rights to those skills.
> theoretical education only makes sense in higher-tier institutions, where the prestige can offset the initial disadvantage
That doesn't make any sense. An institution will hire as per their needs: higher-tier, lower-tier, mid-tier all have their unique needs.
The reason why higher-tier institutions can afford to hire advanced degree graduates is because lower-tier, mid-tier simply cannot afford that baggage: highly credentialed individuals often won't work on tasks they consider "beneath them", the actual "plumbing" of engineering, which makes them useless to lean, high-velocity teams - which is precisely the point the founders at start ups make.
I think you're conflating multiple issues together altough you're not alone in the population that strongly advocates for education as a proxy for ability. They have nothing to do with each other and have a lose correlation.
Here's a thought experiment:
- Kid A (Rural Finland): Has internet access, a Claude Code subscription, and a singular obsession. They spend their time learning how to write a provably correct, memory-safe OS kernel. Through sheer, unhindered iteration and thousands of failures, they master thread management, hardware concurrency, durable scheduling, and modern CPU architecture.
- Kid B (New York City): Attends a prestigious prep school. Their entire mental bandwidth is consumed by a generalized conveyor belt: memorizing 100 years of Eupropean and U.S. history, computing the molality of an acid drop, identifying stereocenters in organic chemistry, and writing essays on what Socrates said to Glaucon. They barely touch operating systems because there is no time left after juggling DBMS trivia, analog design, and quantum physics to satisfy a rubric.
Which person would you listen to or hire when it comes to building an OS?
Now your argument might then pivot to "but what if instead of building an OS, I wanted someone to do algoritmic trading on Wall Street? Ha! I got you there! Kid A knows nothing about that, they only know how to write a kernel!"
Kid A doesn't care. Their singular obsession is OSes and they are world class at it. Their brains aren't clouded by molality and stereocenters - an OS kernel certainly doesn't care about them. If in the future, we had a CPU that worked off organic chemistry, maybe he will learn about it or maybe not and there will be Kid C in the unlikely case Kid A failed to build a trading engine from scratch, fail 1000 times, and learn the math on the fly just like they did for the kernel.
---
[1] Google's Internal Data (Laszlo Bock): In 2013, Google's SVP of People Operations publicly revealed that GPAs, test scores, and university prestige had zero correlation with employee success. They explicitly began hiring more candidates with no college degrees because work-sample tests were far better predictors of impact.
[2] The Dale & Krueger Study (NBER): A famous National Bureau of Economic Research study found that students who were accepted to Ivy League schools but chose to attend less prestigious state schools earned just as much money later in life. It proved that the ambition of the student drives success, not the prestige of the institution.
[3] The Schmidt & Hunter Meta-Analysis: The gold standard in industrial psychology (examining 85 years of hiring data) found that "years of education" has a dismal 0.10 correlation with job performance, whereas "work sample tests" and "cognitive ability" are highly predictive.
The actual Finnish kid who wrote a somewhat successful OS kernel was born in Helsinki, to an upper middle class family with a communist father. He chose to study in the most prestigious university in the country, largely due to his family ties. Because the CS program was theoretical, it didn't take that much of his time. He could enjoy student life and also pursue his other interests. The OS kernel emerged from the latter.
That's what education in a research university is like, when you actually take advantage of it. You don't have to focus on whatever seems useful and practical in the short term, you don't have to compete against your peers, and you don't have to overwork yourself. Instead, you have a socially acceptable opportunity to study whatever interests you, and you hopefully gain the attitude that you can learn whatever you need to learn.
I hope this is a joke. Linux runs on toasters, rockets, phones, tablets, TVs, and systems I can't even think of right now.
There's a popular public debate about an OS professor, whose book is studied all over the world (and I had to study from it as well) publicly berating him for his monolithic kernel design over a proper modular kernel. We are lucky he stood his ground regardless.
> Instead, you have a socially acceptable opportunity to study whatever interests you, and you hopefully gain the attitude that you can learn whatever you need to learn
This isn't unique to a research university. My retired neighbor, who has no academic background except the education he got as a fighter pilot in the USAF, studies whatever interests him all the time. In March, he reached out to me to learn about Wi-Fi HaLow so we could broadcast internet to his remote security cameras, and I believe he's teaching himself YOLO object recognition in August to detect people from those feeds.
Humans have always had the ability to learn whatever they needed to learn. Fire, the wheel, and farming happened thousands of years ago, long before the first research university was established.
In fact, when I visit a modern research university, I find it claustrophobic. Unless you are part of the groupthink and do research on pre-approved subjects, you're very likely to be ostracized until you either "fall back in line" or quit.
Cursor's "Auto", to me, used to mean "cheap flat rate, don't worry which model ran". That assumption is now a financial mishap waiting to happen as of today: 2026-07-22.
Cursor just shipped "Cursor Router" [1].
Auto in the model picker is now three modes, no longer one:
- Intelligence - frontier quality
- Balance - "daily driver" frontier quality
- Cost - the previous Auto routing, with fixed per-token prices (per Colin on the forum [2])
Balance and Intelligence now bills at whatever model the router actually picks and *draws from your API usage tier*.
So if you're still on "Auto" and see "GPT-5.6 Sol (Auto Balanced)" billed to your account under On-Demand, that's now *expected*: while you didn't pick Sol manually, Cursor's "Auto Balance" (which I believe is the default?) did, at Sol's API rates and billed you for it.
It's possible I was the only person caught off-guard with this change - a banner/toast/popup would really have been nice but in case you're caught off-guard too, now you know.
Consider it yes, However having experience in this ... not really. For now there is a file called Decisions.md in the repo that is my "notes to self" if you will about where and what I need to do.
> What do people use uncensored Grok for (like, real use cases) that they can't or won't use other LLMs for? Literally the only thing I can think of is generating bad porn of unconsenting people
untrue. There's a full thread about it: https://news.ycombinator.com/item?id=48837162 - but as much as I love Claude products, nothing's more aggravating than it refusing to help me diagnose a stack trace because it "violates Anthopic policy".
> anyone who is willing to build on stolen property just so that we can accelerate enshittification and damage the environment
Others have made solid arguments(Stackoverflow, open weight models, and fully open source models) - but I encourage you to study the ecosystem post War II and the 1970's Silicon Valley, especially how the semiconductor companies "innovated".
The tiny little MOSFETs you're employing to read this are all built on stolen IP.
As someone famous said, good artists copy; great artists steal.
> The tiny little MOSFETs you're employing to read this are all built on stolen IP.
The alternative to an LLM typing code is a human typing the code. What is the alternative to microchips?
> As someone famous said, good artists copy; great artists steal.
And people who could not be further from artists, or even art enjoyers, think stealing makes them artists, too. Because they're that far removed from art, or any grace, really.
Which doesn't go against OP or the project, which I find delightful. Although I generally share many reservations of "AI critics", I'm also a starved: if it's snappier and uses less resources than something humans coded, come right in! At least if it's a neat thing like this seems to be, and not some sprawling trojan horse with code that "works" but looks terrible... machine optimized stuff that is opaque to and not made for humans need not apply.
I'd bet 99% of overall token usage has nothing lasting to show for it, and of the 1% that actually compound into anything 99% are nothing like this. So just like a broken clock can be right twice a day, a super correct soldier of Butler can hold fire twice a day, no?
There are different ways to answer this. One person gave a literal answer - Vacuum Tubes. My, alternate answer is, humans. Computers are machines that automate human processes. If we didn't have a digital domain, we would just be doing what we were doing before it, printing and writing.
> I'd bet 99% of overall token usage has nothing lasting to show for it, and of the 1% that actually compound into anything 99% are nothing like this
What you're betting with?
LLMs have changed to game on velocity of knowledge work. The future has changed and "nothing lasting to show for it" is a very limited take on the matter.
The idea that "the future is changed now, so you can't change it again" is the language of abuse, not the language of wisdom. We'll just change it again by draconically punishing thieves. Then we can have the best of both worlds: cool tech, and no Eloi/Morlock bs.
Anyway, I simply meant that most vibe coded stuff doesn't seem to get maintained, and that most token usage doesn't produce any artifacts to begin with.
> "the future is changed now, so you can't change it again"
that's not what I said.
As technologists, we repeatedly change the future - that's what makes all the pain, sweat and tears worth it. 35 years ago, when I was getting serious about computers, buying them and encouraging others to use them, people would shake their fingers at me, saying how these bulky TVs that you can press buttons at were a passing fad, that it was an absolute waste of everyone's time even thinking one would sit in front of a TV and press at buttons for more than an hour a day, that there was no reasonable way it would translate into anything but extremely niche entertainment. They would lecture me on how I was a kid who didn't know any better, that I was wasting my time out of ignorance and that I should listen to the adults who had decades of life experience and focus my time and continue studying organic chemistry and become a doctor because who didn't need a doctor?
> most vibe coded stuff doesn't seem to get maintained
even if this was true, cost of code consumption and generation has fallen so much, that anyone in a developed country can point their favorite agent to the
"unmaintained vibe coded stuff" and then maintain it to their liking or rewrite it from scratch the way they wanted it to.
> most token usage doesn't produce any artifacts to begin with
Your previous statement contradicts "doesn't produce any artifacts" directly. Given what you have told me, the shallowest but congruent argument that can be made is "most token usage produces unmaintained vibe coded stuff".
It looks like you don't necessarily like what LLMs are providing to society and I can see why one would like to hold that opinion. I don't agree with that at all, because it's literally untrue given the insane demand - both from an everyday Joe and from corps the world over. No one's burning $20+/mo every month of their own money in this economy just because they are not getting anything of value.
My personal AI spend is $350/mo and it has been that for the last year. My blockers are gone. Projects that had been a distant thought, only cosidered during a flight, a wait at an airport or for a bus or Uber is finished in a weekend. My QoL and those of my family has improved so much, that I am really grateful about the times I live in.
My 100 year old grandfather struggles to communicate and remember things. I cannot expect to give him a tablet and have him use it. We have spent years looking for apps that we could use to make his life better. None of us are mobile app devs. Solved in a week.
I had random headaches when I woke up really early, ever since I was a teenager. Tens of thousands of dollars spent chasing lab tests and doctors, no solution. Resolved in a month.
The most I can do is encourage you to set aside your current stance, and just for one day, consider what if you could use LLMs to improve your life. What would you do?
> even if this was true, cost of code consumption and generation has fallen so much, that anyone in a developed country can point their favorite agent to the "unmaintained vibe coded stuff" and then maintain it to their liking or rewrite it from scratch the way they wanted it to.
Yes, and? In context, my point was that as far as vibe coded stuff (that I saw so far, and that gets posted on HN), this is head and shoulders above the usual goes. This general evangelism of LLM that isn't even aware of the context is really grating.
> It looks like you don't necessarily like what LLMs are providing to society
I hate what they took. If you steal from Meta, they sue you. That is all there is to say.
> The most I can do is encourage you to set aside your current stance, and just for one day, consider what if you could use LLMs to improve your life. What would you do?
Do you not understand that ethics can be a value in and of itself? I'm not ethical to interact with trinkets, I interact with trinkets (with gimmicks, with toys, with product, with dust) to be ethical. The best you can have you were born with: integrity. How much of that we can preserve is the only challenge in life I take seriously. You're asking someone who dines like a king every second what they would cook if they were "allowed" shit as an ingredient.
I have no "qualms" toying with LLM. I did, I got bored, so now I don't. When it comes to generating "art", I actually do refrain from using it even when it could be useful, because yes, it's generally theft. Wake me up when there are LLM fully trained on licensed content, and in the meantime, accept a no.
> A diverse market full of choices keeps it from becoming the browser wars all over again.
This is a great analogy but I worry you might be implying something I don't agree with but you didn't explicitly say what I'm worried about, so let me call it out:
Microsoft played a dirty game with I.E, but they are in the dirty game business. It wasn't only I.E, it was their OS, Office suite and everything else they do business in.
Google Chrome took advantage of that dirty game and now you have the Chromium engine that powers a lot of browserlike frameworks.
No one born in the LLM age even knows what I.E means or stands for, as it should be - a horribly designed, poorly working product foisted upon users via the Windows distribution system - a dishonorable product from an ethically corrupt company forever lost in history, right alongside Clippy and DCOM.
OTOH, I am glad that Microsoft played a dirty game with I.E and didn't just stop playing dirty there - they jacked up the price of Windows if an OEM even dared to bundle in Netscape Navigator instead - who knows, if they hadn't done that, there wouldn't have been a Google or Apple. We would all be using Windows and Windows Search and Windows Phone.
And without Google, we might not have had the modern LLM as we know it. We would have had some trashy Windows Autocomplete Copilot Clippy. Ugh!
"No one born in the LLM age even knows what I.E means or stands for, as it should be - a horribly designed, poorly working product"
As one of my first jobs involved getting a website to work with IE6 I surely hated it, but when it came out, it seemed to have pushed the web technologies in general.
The problem was not the browser technology, but microsoft abusing it's monopoly to don't give a shit about (open) web standards.
I still have no clue what you're implying that I was implying? Reading too much into someone's words is definitely a thing.
There used to be 4+ major browser engines. Now there's only two, and Google owns almost complete share & following standards? Google prefers their monoculture.
More diversity in browser engines was a good thing & standards were lovely.
Not everything has to be a slug-fest between #1 & #2.
& personally I'm glad there's Grok & Gemini to keep Anthropic & OpenAI on their toes even more than the competitive band of open weights models already do.
The corporate models tend to be defensive and sound like HR has approved every word of the script. Not a fan.
My hypothesis is that all the top providers realize that, lacking vendor lock in, all SOTA models in a year or so's time will be similar in capability. Also, open weights models are continuing to catch up in a year's time, sometimes less.
So they are trying to lure you in with differentiating, superior capabilities into their proprietary, non-open, non-standard agent harness.
It's the Hotel California playbook: These amazing capabilities are to attract you like moths to a flame and keep you warm and alive around the flame but waterboard and shock you if you attempt to move away from it. Like AWS Egress charges.
Nein. Neither Pfizer nor Citadel are going to use a hosted model - their whole business is IP. They will invest in self hosting models. Self hosting is becoming a big deal in the industries you're talking about.
Why?
Neither Pfizer nor Citadel wants to pay a supplier/vendor to build a core capability that the supplier/vendor will turn around and sell to their competition.