The point is that it is inefficient. Others achieved similar results much cheaper, meaning they can go much further. Compute is important, but model architecture and compute methods still outweigh it.
To clarify, LLMs will lower the entry barrier for non-programmers to test new ideas (which is good). Later, those successful projects will require development by professional developers. I suspect that many of these projects will have lower quality than regular proof-of-concept projects today.
This, plus in several years we will observe many MVPs generated (partially) by LLMs that will require fixing and further development. LLMs barely produce 1k lines of coherent code, riddled with bugs. Benchmarks are improving, but there is no breakthrough in handling real, large codebases. I have no idea about the future job market, but the quality of the "material" we work with will be quite poor.
20 years ago, one of my professors pointed out during a review that "benchmark is not research." Now I see literally everywhere benchmarks with very basic conclusions. I'm not criticizing this particular work, but such publications are low-effort unless they point to something new.
Ghost jobs on one side, ghost applications on the other. Some people will just send automated applications everywhere, every day, and check for responses. That leads to ghost responses, and the cycle continues.
Second that. GDPR actually made those aspects clear and never caused a headache during implementations I've seen or participated in (more like a checkbox on a list). When I see any complaints, then it's clear some iffy user sniffing is happening.
That's the main point. No one buys software, they buy solutions. Accounting is a good example. I use a SaaS solution, but it doesn't matter because I could also take all my invoices to an accountant, and the effect would be the same.
Also, mixing open licenses into business doesn't usually make sense.
I also think that mass source scraping for ML/AI training will make businesses less likely to participate in open source.
Since I scroll quite a lot, I prefer to have a better experience with this very basic and common activity. Apple claims to provide the best experience, while it clearly does not.
Well, the CS program differs slightly between polytechnics and universities. The former emphasizes practical engineering, such as networking details, while the latter focuses more on theory, such as algorithm correctness proofs. I think our job requires a lot of engineering due to reliability, performance, compatibility, communicating design, etc.
Yes, one very common use case where Postgres does not scale well is analytics. Snowflake, Vertica, and ClickHouse work much better. I say this after working on several projects where development teams hit a wall with row storage databases. Still, PostgreSQL is great, it's my default DB as well.
Enshittification is actually a hilarious process to watch when you know it's happening and have a way to escape it by switching to an alternative. I escaped from HP, Windows, Reddit, changed banks, and moved to a bunch of other products, sites, and services.
This is another take on defunctionalization. You create a model of execution but do not execute it. I.e., return or queue a value of type Send, and do not execute "send". The execution is separate and actually deals with "real-world" side effects. The execution can be done by sync/async/transformed to monads, it doesn't matter.
Nokia not only missed what users wanted but also how developers would like to work. Symbian had a strange and easy-to-misuse API, poor manuals, and app deployments were tied to telecom operators as there was no app store. I remember my ramp-up as a developer with the Apple ecosystem after working with Symbian. It wasn't evolution; it was like a leap to the next century. The documentation was excellent, the APIs were clear and logical, the code made sense, and the system was familiar. And then the App Store arrived with all the implications, such as a sudden mass market and less-obvious features like crash reports with stack traces from users. Apple was on an entirely different level than Nokia, and literally not a single developer wanted to touch Symbian anymore. It was effectively dead, even with millions of devices still in use, but it was obvious. Today I see a similar story unfolding, much more slowly, but the same thing is happening with Microsoft Windows.
Absolutely agree. As most projects are small or medium in size, PostgreSQL is a good default choice. Problems arise when the problem does not fit the solution or tool. For example, using SQLite for high-volume transactions and then complaining that it doesn't work.
There is no one-size-fits-all solution in the SQL world, and using SQLite in such a scenario is not a good idea. SQLite is an embedded database for embedded use. Even PostgreSQL is not universal, as many people assume. For use cases like analytics, there are many better-suited database engines.
It depends on the company. One company I worked for had a policy of "Just buy what you need that makes your work easier: books, licenses, and even hardware. Management time spent on decisions is too expensive, so we default to instant approval." Did developers go on a buying spree? No, the same amount of stuff was bought as before the policy. And it did make life easier.
It's more profitable for Amazon in the long term, as they can potentially eliminate competition. Amazon's use of "open-source advocacy" is questionable, and I wonder if anyone believes it.
Exactly this. Most of the tests reimplement the implementation using mocks. Such tests are useless, as they always prove the code is correct. Worse, such tests make refactoring much slower. On a low level, only black-box interface tests make sense, and on a high level, use scenario testing. The implementation has to be tested indirectly, otherwise, it leaks.
It appears that my previous comment lacked clarity. I assumed the inherent absurdity and illogicality of the statement would be self-evident. Dismissing François's opinion, a seasoned professional with extensive expertise in model development and the creator of ARC, solely because he asserts that LLMs are far from achieving AGI and using publication count as the metric for evaluating him is not a well-reasoned argument. While LLMs and transformers are indeed remarkable achievements and will undoubtedly reveal more properties, they have yet to exhibit any true signs of "intelligence."
Short whiteboard interviews have become LeetCode/take-home assignment/one vs whole-team-from-every-department interview marathons. So instead of those funny pictures, we'll have Rorschach tests followed by polygraphs soon.