> Is the lock structuring here really deadlock safe? The model will tell you with complete confidence its code is perfect
Fully agree, in fact, this has literally happened to me a week ago -- ChatGPT was confidently incorrect about its simple lock structure for my multithreaded C++ program, and wrote paragraphs upon paragraphs about how it works, until I pressed it twice about a (real) possibility of some operations deadlocking, and then it folded.
> Every time a major announcement comes out saying so-and-so model is now a triple Ph.D programming triathlon winner, I try using it. Every time it’s the same - super fast code generation, until suddenly staggering hallucinations.
As an university assistant professor trying to keep up with AI while doing research/teaching as before, this also happens to me and I am dismayed by that. I am certain there are models out there that can solve IMO and generate research-grade papers, but the ones I can get easy access to as a customer routinely mess up stuff, including:
* Adding extra simplifications to a given combinatorial optimization problem, so that its dynamic programming approach works.
* Claiming some inequality is true but upon reflection it derived A >= B from A <= C and C <= B.
(This is all ChatGPT 5, thinking mode.)
You could fairly counterclaim that I need to get more funding (tough) or invest much more of my time and energy to get access to models closer to what Terrence Tao and other top people trying to apply AI in CS theory are currently using. But at least the models cheap enough for me to get access as a private person are not on par with what the same companies claim to achieve.
Hello, TCS assistant professor here: he is legitimately respected among his peers.
Of course, because I am a selfish person, I'd say I appreciate most his work on convex body chasing (see "Competitively chasing convex bodies" on the Wikipedia link), because it follows up on some of my work.
Objectively, you should check his conference submission record, it will be a huge number of A*/A CORE rank conferences, which means the best possible in TCS. Or the prizes section on Wikipedia.
I don't deny that his output is highly valued among AI researchers.
Provocative as my question may be, the point I wanted to make is that his most highly cited paper that I already mentioned is suspiciously very in line with the OpenAI narrative. I doubt if any of his GPT research is really independent. With great salary comes great responsibility.
> Realistically there's no reason government can't use open source software and open formats especially.
> Last time I had to fill out a government form in Canada (...)
Without any evidence, let me argue why maybe it shouldn't. In the past, a common opinion that I have heard is that open source is more secure because all the code is out in the open.
The recent xzutils backdoor attempt [1] kind of led me to believe it's not really true, it's only true if many good-actor eyeballs, which are willing to donate their time for public benefit, are on the code.
Almost all of the government's code that I interact with are web apps that are potential targets of foreign adversaries -- tax filing web apps, prescription + vaccination scheduling web apps, family benefit applications, and more. (This is not in Czechia, but close.)
Now, would I want to read that web app code? Not at all, I couldn't care less about it. However, foreign adversaries would love to immediately start analyzing it. Extracting the entire country's health data or tax data would be a goldmine.
And even though there probably are several people actively paid to maintain security of these systems, I feel that the foreign adversarial agents would be much more motivated (and better paid) than government employees/software developers.
You could make a opt-out for national-security purposes for the code, but I feel almost all the code a government works on would have such an impact when compromised.
(Disclaimer: I am a huge supporter of open source in general, contributed to the Linux ecosystem in the past and in my current job as an academic, almost everything I do is available out in the open in some way or another.)
Microsoft has literally been hacked multiple times by Russia in the last few years. Our government lost hundreds of thousands of CRA (tax agency) credentials to hackers and had to lock millions of accounts. Other agencies have also been breached.
Meanwhile the XZ backdoor was found in Sid, Arch and pre-releases of Fedora and openSuse. It never actually made it into any numbered release of Fedora, openSuse, Ubuntu, Debian, Red Hat or Suse distro. It's actually a pretty big win and the system worked as intended.
Open source and Linux are doing just fine security-wise.
Also, none of this has anything to do with using offline tools like a word processor to make documents.
>Meanwhile the XZ backdoor was found in Sid, Arch and pre-releases of Fedora and openSuse. It never actually made it into any numbered release of Fedora, openSuse, Ubuntu, Debian, Red Hat or Suse distro. It's actually a pretty big win and the system worked as intended.
I would maybe not go quite that far. That it got caught was mostly a confluence of lucky breaks and accidents. The second version of the exploit would likely have not been detected if not for the fact that the first version of the exploit had a couple of programming mistakes that attracted some attention to itself.
The entire thesis behind the open source security model is to have lots of eyes on the code/program, since more eyes = more likelihood of catching it. Even if you say it's accidental, let's say the odds of catching it are 0.00001. Repeat that enough times and you get 1.
It was caught before any distro released with it. The system worked.
If one of the Debian or Fedora developers had immediately caught on to what they were looking at when their attention was drawn to it by the failures, I would say the system worked. It's certainly true that open source saved the day here, but that's maybe different from saying "the system" worked. It easily could have gone unnoticed, or been noticed a few weeks later.
The xz backdoor was caught before anyone used it. This is typical of open source backdoors, but atypical of proprietary ones. History is full of proprietary software with backdoors which were discovered after years or decades of being actively used. Lotus notes, RSA corporation, Cisco routers, Juniper switches, Huawei everything.
We have more or less immutable history of every change leading to every release of open source software. Any backdoors you previously created under an identity could burn that identity forever. That history is not available for proprietary software. If someone adds a backdoor in proprietary software for two years and then removes it in later versions, it's totally likely it'll never be noticed.
Thinking that open source software is at greater risk of being backdoored is akin to thinking most trees in the world grow along the road, just because you drive everywhere and have never been inside a forest.
Every country already has a special government agency that deals with keeping stuff protected. In fact you tend to think the people who know most about this are in government, don't you?
And it's not like there haven't been vulnerabilities found in proprietary software, despite them paying people to keep things safe.
Crowdstrike is a recent example that comes to mind. I don't see how paying for CrowdStrike made it more secure or reliable.
I would also argue that you could take all the $$ paying for proprietary software and contribute it to people who are making the open source software, making the reliance on "free" eyeballs less of an issue.
> I've worked in academia - and yes, that has been in multiple countries in the EU - I've never had to utter a single word in something other than English in the workplace.
I have the same trajectory as you -- multiple countries in the EU, working in academia -- but different experiences for sure. Or at least a mixed bag.
Let me list them in order of how much English sufficed:
1. The Netherlands -- common knowledge is that their English is top notch and anecdotally it was the case as well, I also got by purely with English.
2. Germany -- their English is also good but I needed German in edge cases. One edge case was finding an apartment (not speaking German simply pushed you down the list of candidates, even with a full time job in academia). Another one were university rules and announcements; not every email was in English, but arguably easy to get by with modern translation tools.
3. Czechia & Poland -- English is good among the professors but the percentage of locals at the university level is so high that most internal meetings, announcements, local seminars take place in the local language. In my experience, non-faculty university staff (department secretaries, payroll, entrance security) usually strongly dislike speaking English or outright do not speak it at all.
---
I've omitted some more cases where local languages are required. If you live in a country, you will eventually interact with the healthcare sector, where the language experience will likely mimic the experience at the workplace (for the countries above, it would be in the same order for the healthcare sector).
Another case is government bureaucracy. For most of the EU countries I've been to, the official language of the country is their local language and only their local language. This means that government employees are not required to speak any other language other than the official one to you, plus you might be required to fill in forms and communicate in the official language if you want to talk to them.
In my experience, the helpful/good ones may try to communicate with you in English but if you need something from them or if the bureaucrat had a bad day, you better start talking in the official language.
> Another case is government bureaucracy. For most of the EU countries I've been to, the official language of the country is their local language and only their local language. This means that government employees are not required to speak any other language other than the official one to you, plus you might be required to fill in forms and communicate in the official language if you want to talk to them.
This is true, and something I have indeed experienced. However, this is likely true for _any_ country where English is not the official language, not just those in the EU. Besides, understanding bureaucratic lingo is not just a matter of pure linguistics. Governmental concepts rarely translate 1:1 to another nation, even those with the same official language. If you migrate to another country, part and parcel of the experience is that you _must_ contend with bureaucratic principles, rules and institutes with which you are not familiar. There is no escaping that.
That said, at least here in the Netherlands, there is certainly a movement to provide more and more governmental information in English as well. I'm not going to dox myself, but for example my muni's English website looks nigh-identical to the Dutch one.
The problem is social life and informal discussions. I France or Germany you cannot have a normal life without a fairly good knowledge of the language.
It reminds me of the fact that the "Fake Mr Beast giveaway" ads that even raised some attention here on Hacker News [1] a while ago are still around. In fact, I have seen one yesterday. Those must have been flagged as impersonation and scam thousands of times by many people, including me personally, and Youtube finds them perfectly fine.
After that episode, where I tried myself to get rid of them, I am much more convinced that Youtube is fine with all but the worst scammers, and don't buy any of the "they're just low on manpower" arguments anymore.
Last year for weeks I was seeing these fake Mario games on YouTube ads. And it wasn’t a Mario feel-like. They called it “Super Mario” and had whole asset rips in the graphics.
Nintendo lawyers were surely all over that. I’m shocked it took so long to get them removed.
> Hi! I'm a PhD student studying computer science at Rice University.
This means that we are on the same career path (I am currently an assistant professor in theoretical CS in Europe). I wish you of course best of luck!
Here is the harshest truth about teaching I learned during my PhD:
If you are focusing on teaching too much, you are setting yourself up for failure.
This sounds cruel, and in fact I am much like you, I love teaching and I love self-improvement and it is quite easy for me to invest time into my teaching prep, presentation, and more and see measurable results in class quality and usually also student feedback.
However, at least in my neck of the woods (i.e. Europe), almost all gates and gatekeepers for you as a PhD student, and later postdoc, are checking your research. At some places they really do expect you to have K publications in the top 3 CS conferences or you will not be considered at all -- and it seems these thresholds are only getting higher. Here I mean for example invitation-only workshops, postdoc positions with top advisors, and later also permanent positions.
On the other hand, if you are a talented scientist, they usually only care that your teaching skills are at the bare minimum -- have you taught something? Yes? Great.
Now orator/presentation skills are critical and presenting a coherent lecture plan might be useful for a final presentation at an interview for a permanent position. But even there, it is more about you knowing what you want to teach and how it complements the department than about your past achievements (i.e., how much you have put in a course previously).
My PhD advisor usually said that he likes to dig into teaching when research is not going well. I agree with that -- teaching really is fulfilling to me and I love to improve my class and see people happy with it, and research is all about global ranking (which is tough on anyone's psyche) and generating progress which is the fun part but sometimes takes a long time. However, at your stage of your career, the research really can't go slow.
---
PS: If the author reads this, since it is a self-post, your class sounds really nice and it is actually one I would have loved to attend. My research is in online algorithms -- a field which you can rephrase as seeing some theoretical problems as two player games between a solver and an adversary -- and among other things I would like to consider utilizing all the techniques of chess solvers (which cannot evaluate the game fully, but "almost") and transfer it to other areas of online algorithms.
Just as a counterpoint: this very much depends. I probably spent at least a year (probably more) of my PhD (in Europe) just teaching a class I built up from the ground up myself. I barely got any research done the first year I gave that class, and every subsequent year it still took a large chunk of my time. It's part of the reason I spent a total of 7 years doing a PhD (which is long, considering I already had an MSc), during 5 of which I taught my class, and grew it from 10 students in the first year to 200 in my last. But I don't consider that time wasted. I had a blast and found that teaching helped me understand the fundamentals of my fields at an extremely deep level that I'd never reached otherwise. It didn't improve my research output, but I feel that the soft skills and understanding of fundamentals was a real advantage. My future career also didn't suffer, I'm now working as researcher at a FAANG AI lab.
> I had a blast and found that teaching helped me understand the fundamentals of my fields at an extremely deep level that I'd never reached otherwise
You spent 5 years teaching a class that, judging from your words, you probably prepared and improved very thoroughly. That is a lot of hours of work. Are you sure if you devoted all those hours to reading textbooks, papers, doing experiments, etc. on your field, you wouldn't have achieved an even deeper understanding?
Maybe yes, but if so, I honestly think you're in a minority. As an academic myself, I like teaching and I do learn things from it, but it's far from the most efficient way to learn a scientific field. If I had a pure research position I'm pretty sure that my research productivity would be better.
> If you are focusing on teaching too much, you are setting yourself up for failure.
This is good advice. And this is true even once you become a professor. All time spent on teaching will go against your career progression. Even if you're tenured and don't care about promotion, you'll feel like an imposter in your department if you're not somewhat competitive research wise.
Generally speaking, there's no recognition in teaching in general, and at university level it's often not even considered as a job by itself.
Maybe it's different in Asia, but that was my experience in the western countries where I worked.
> However, at least in my neck of the woods (i.e. Europe), almost all gates and gatekeepers for you as a PhD student, and later postdoc, are checking your research.
While I'm also in Europe, my bet is that this is universal and won't change in the foreseeable future.
The reason is that teaching is practically impossible to evaluate. How do you quantitatively measure which professors provide high-quality teaching? By grades? No, easiest course wins. By employability? No, it depends a lot on the field, a philosophy professor can be amazing but that won't create jobs in philosophy. Student polls? Correlation with actual quality is really weak, and I say this as someone who has good polls - there is a strong influence of difficulty as well as the subject itself (a CS student will almost always prefer programming to physics, and it's not the physics professor's fault), apart from gender bias.
In my country they try to give an equal weight to teaching equally with respect to research in applications for positiosn and tenure, but since there is no realistic metric, the bulk of the score ends up being about "years teaching" or "number of hours taught" which is the only objective number that they can come up with. So it becomes basically a seniority factor and since your seniority is what it is and preparing high-quality lectures won't give you more hours or years, the outcome is still that focusing too much on teaching is bad for your career.
There are different types of universities. While R1 institutions are more focused on research than teaching, there are smaller liberal arts universities which revolve around the undergraduate student experience. These universities still have research expectations as part of tenure and promotion, but faculty aren’t required to crank out research publications. Teaching is hugely important at these schools, both during the hiring process and when evaluating candidates for tenure and promotion.
I have been fortunate enough to work at such a university for the past 20 years. We have a deep endowment, small class sizes, and extensive support for our faculty research projects. Undergraduates at our school are often engaged in research projects as well.
For me, this is like an academic utopia: a blend of teaching and research with a primary focus on teaching. There are many other universities like mine.
Yes, I'm fully aware of the fact that teaching isn't really a priority in academia - for that reason, I probably won't be reviving my class in the near future. I really do like teaching, but it doesn't get me much closer to any of my current goals.
> Isn't the problem de facto solved by matchmaking? The player that aims better will quickly win more and be elevated to the level of opponents on par with them.
Matchmaking decreases the odds you meet a cheater for low rank players, and significantly increases it for higher rank players -- and since there's fewer of them due to the Bell curve, they are going to feel the cheaters that much more.
If you just rely on rank and not on anti-cheat efforts, you'd be just destroying one of the loyal cores of the playerbase, one which is also quite vocal online.
From my personal experience of thousands of hours in competitive FPS shooters on PC, there is no point in ranking where playing against a cheater becomes fair or fun.
modern matchmaking includes streamers that pay for accounts that have purposefully lost repeatedly so they've tanked their rank so they can stream themselves pubstomping noobs.
Back in the day an obvious cheater would get booted from the server, nowadays, they literally record themselves doing it and nothing happens because it brings in an incoming to the company.
> Too often, grad school applicants are just kids that have overachieved in academic settings and think to themselves “I’ve been good at school my whole life, why not just do school forever?
Anecdotally, I did not observe this during my PhD studies (theorerical CS) in Central Europe. I think this might be due to the separate 3 year Bachelor track, then a 2 year Master track, and only then 4 year PhD studies.
Sure enough, a lot of applicants faced tough career decisions after graduating, but whoever started the PhD usually knew what research is about and that it's going to be work first and foremost, not just "more school".
Quickly skimming it, I found no evidence of what the future actually held, from Wikipedia [1]:
> In 1981, Pizza Time Theatre went public; they lost $15 million in 1983. By early 1984, Bushnell's debts were insurmountable, resulting in the filing of Chapter 11 bankruptcy for Pizza Time Theatre Inc. on March 28, 1984.
Rapid expansion seemed to serve the owners goals (from Nolan Bushnell’s Wikipedia page. He also started Atari.)
> It had been created by Bushnell, originally as a place where kids could go and eat pizza and play video games, which would therefore function as a distribution channel for Atari games.
Little known fact, the idea came to Bushnell when he was still running Atari. He was waiting for pizza and realized these bored people would pay to pay games while waiting. [1]
Bushnell was frustrated that Atari only got money on the original sale of a coin-op machine, while the operators got continuous revenue from the machine. He figured Atari could get in on the operations game by running a restaurant.
Atari built/ran the first Chuck E Cheese and Atari engineers (and Atari think tank Cyan Engineering) designed the robots. Restaurant opening party was private for Atari staff.
The mascot was originally going to be called Rick Rat, until marketing suggested that associating a rat with the restaurant wasn't the best idea.
When Warner Communications bought out Atari, they didn't understand it or want it, so Bushnell bought it out from them and continued running it.
Yeah instant 2x headcount growth is a pretty blaring warning. Good thing we learned the lesson well in the 80s and haven't repeated it since.
And these lessons are incorporated into business programs world wide and taught to each new generation so that they may permeate the business landscape and prevent unnecessary hardships.
This is why I advocate advanced degrees in business for folks looking to enter leadership positions.
OK, but… that’s a store selling anything and everything under the sun tapping into a previous untouched market. This is an animatronic rat for kids. Context is key.
They pulled in more revenue than the average McDonald's and 3x more than Pizza Hut.
It does seem like the concept has always struggled to be cost-efficient but I don't know that rapid growth was a bad indicator on its own.
It's also probably worth keeping in mind that some of those restaurants may have been franchised. When a franchisor sells a franchise license to a local franchisee, the local franchisee bears most of the risk.
Pretty sure McDonald's and many other chains have seen that kind of growth at one point or another. Going from 20 to 40 or 100 to 200, in a year is not necessarily an indication of a problem.
Past etymology isn't really relevant to how a word is actually used today. There are plenty of awful slurs that were once everyday words used offensively.
I am currently wearing a "don't be a Richard" shirt which is one of my home improvement shirts. 2 different checkout folks at 2 different hardware stores were calling over coworkers because they both apparently had managers off today named Richard who were in fact "Richards".
Lol now I’m curious what is your other home improvement shirt. I could use an official way to convey to people that I may look like I’m just puttering around but I am actually on a diy mission.
oh by diy shirts I mean all the shirts that I have oil stains on or paint... that I wear when I just don't give a f or hwne I know I'm going to get more dirty. I have one that is an old mozzila appu shirt... with a sweet gtradient on it. The gradient came from me usinga circular slot cutter on a mill that tossed metal and oil at my midline and taperd off as it fell.
Speaking as an (assistant) professor in theoretical CS, I see there are many bad approaches in the original post and mentioned in this discussion, but I strongly disagree with:
> There are very simple and effective ways to teach your class that can't be cheated with AI. These professors are simply lazy and uncreative.
I can attest to the following problems to good homework creation, from my own experiences playing with ChatGPT and teaching:
1. If you want to give a very illustrative yet easy theoretical exercise in algorithm design, one that computer scientists have solved over and over in the last few decades and which furthers your understanding, there are very likely solutions online and ChatGPT will give you the solution with very high probability.
2. If you create your own dataset and want the students to implement some algorithm and create a simple plot/discussion from the results, it will be very hard to distinguish a "student solved it on their own, but they did not invest too much time into it" submissions from ChatGPT submissions produced by a couple of queries.
3. Switching to oral presentations is hard to scale (as others attest) and also does not resolve much, because some students are perfectly okay with being handed a solution from somewhere (colleague, ChatGPT), not understanding it very well, and yet presenting it. Failing these students likely leads to overly difficult classes.
4. In-classroom exams without a computer work best, but they also do not scale very well (a lot of prep/correction needs to go into them) and some students with bad anxiety management skills, which includes me as a former student, dislike them passionately.
---
As you can see, this topic is quite critical for my profession. The ugly truth is that university professors have only a very limited time allocated in their busy workweeks for teaching, and hence they have to take many shortcuts, including suboptimal homework sheets and limited innovation year-over-year. I also do not allocate as much time for philosophy of teaching/improving teaching skills as I would have liked.
If anyone here has novel ideas how to actually implement "a class that can't be cheated with AI", specifically university CS classes, I am all ears.
> If anyone here has novel ideas how to actually implement "a class that can't be cheated with AI", specifically university CS classes, I am all ears.
May not work for you, but as a CS student our department had the policy if that if you failed the final, you failed the course. The finals were usually structured that rote memorization would earn a C- (depending on course complexity and importance). They were all pencil-and-paper exams.
While cheating was policed, collab was encouraged with the proviso that lab submissions needed to be own-work, and they'd run basic comparisons to make sure that they weren't copies. As a result, the administrivia on finals was longer...but there was a little less concern about the rates of cheating.
> not understanding it very well, and yet presenting it. Failing these students likely leads to overly difficult classes
What does a grade even mean in your classes, if not understanding the subject isn't grounds for failure? Isn't the point of a grade to measure exactly that?
> some students with bad anxiety management skills, which includes me as a former student, dislike them passionately
So what? Nobody likes exams, they make everyone anxious, why is that even considered relevant? Did you not pass through exam halls with hundreds of students in them to get to university in the first place? How did this supposedly non-scalable system scale when you were 16?
> If anyone here has novel ideas
Why are you acting as if this is an unsolved research problem?
Here's a novel idea for you: talk to people who teach 15 year olds and then copy the way they do it. They'll probably tell you to do things you don't personally like doing but if it's really "critical to your profession" as you claim, then that won't matter, will it.
For programming CS exams, you can try doing them in a computer lab with internet disabled. These exams should be allotted time much more than needed as one should not be testing for problem solving under strict time constraints.
FYI its oral exams, not oral presentations. You give the student half an hour to solve a sequence of problems and gauge their thinking skills. Scales to perhaps 30 students at most.
I was not very clear about it, but I was discussing regular semester work, as opposed to final/midterm exams. Think courses that are strongly grounded in theory but need the students to experience the coursework, like Discrete Mathematics or Linear Programming.
"Oral presentations" in my case meant presenting a homework solution to the TA in person, in front of the class, and the TA accepting this solution live (or not).
At least at my university, the responsibility for homework structure and homework sheets lies fully on the lecturer, and the TAs are tasked with grading the homework/projects and leading the exercise sessions.
Oral exams are great if they can be done at scale, and I do use them. Some other teachers (as well as the administration) prefer written exams, as there is a clear proof of work that can be analyzed if grades are disputed.
> I was discussing regular semester work, as opposed to final/midterm exams
In my experience, these are completely useless for any core course as you mention. I have tried stuff like this in my courses, and it doesn't work. In fact, I know some students pay others to make the presentation for them, and coach them on the presentation.
Even for course projects, I have graded meetings with students before the final presentation/report. This helps ensure that they are doing the work themselves rather than depending on others. But yes, this takes a lot of time.
> Some other teachers (as well as the administration) prefer written exams, as there is a clear proof of work that can be analyzed if grades are disputed.
Fully agree, in fact, this has literally happened to me a week ago -- ChatGPT was confidently incorrect about its simple lock structure for my multithreaded C++ program, and wrote paragraphs upon paragraphs about how it works, until I pressed it twice about a (real) possibility of some operations deadlocking, and then it folded.
> Every time a major announcement comes out saying so-and-so model is now a triple Ph.D programming triathlon winner, I try using it. Every time it’s the same - super fast code generation, until suddenly staggering hallucinations.
As an university assistant professor trying to keep up with AI while doing research/teaching as before, this also happens to me and I am dismayed by that. I am certain there are models out there that can solve IMO and generate research-grade papers, but the ones I can get easy access to as a customer routinely mess up stuff, including:
* Adding extra simplifications to a given combinatorial optimization problem, so that its dynamic programming approach works.
* Claiming some inequality is true but upon reflection it derived A >= B from A <= C and C <= B.
(This is all ChatGPT 5, thinking mode.)
You could fairly counterclaim that I need to get more funding (tough) or invest much more of my time and energy to get access to models closer to what Terrence Tao and other top people trying to apply AI in CS theory are currently using. But at least the models cheap enough for me to get access as a private person are not on par with what the same companies claim to achieve.