I do not work in the space at all, but it seems like Cloudflare has been having more network disruptions lately than they used to. To anyone who deals with this sort of thing, is that just recency bias?
It is not. They went about 5 years without one of these, and had a handful over the last 6 months. They're really going to need to figure out what's going wrong and clean up shop.
The featured blog post where one of their senior engineering PMs presented an allegedly "production grade" Matrix implementation, in which authentication was stubbed out as a TODO, says it all really. I'm glad a quarter of the internet is in such responsible hands.
Management thinks AI tools should make everyone 10x as productive, so they're all trying to run lean teams and load up the remaining engineers with all the work. This will end about as well as the great offshoring of the early 2000s.
Wait till you get AI to write unit tests and tell it the test must pass. After a few rounds it will make the test “assert(true)” when the code cant get the test to pass
No joke. In my company we "sabotaged" the AI initiative led by the CTO. We used LLMs to deliver features as requested by the CTO, but we introduced a couple of bugs here and there (intentionally). As a result, the quarter ended up with more time allocated to fix bugs and tons of customer claims. The CTO is now undoing his initiative. We all have now some time more to keep our jobs.
Thats actively malicious. I understand not going out of your way to catch the LLMs' bugs so as to show the folly of the initiative, but actively sabotaging it is legitimately dangerous behavior. Its acting in bad faith. And i say this as someone who would mostly oppose such an initiative myself
I would go so far as to say that you shouldnt be employed in the industry. Malicious actors like you will contribute to an erosion of trust thatll make everything worse
Might be but sometimes you don’t have another choice when employers are enforcing AIs which have no „feeling“ for context of all business processes involved created by human workers in the years before. Those who spent a lot of love and energy for them mostly. And who are now forced to work against an inferior but overpowered workforce.
I dont like it either but its not malicious. The LLM isnt accessing your homeserver, its accessing corporate information. Your employer can order you to be reckless with their information, thats not malicious, its not your information. You should CYA and not do anything illegal even if your asked. But using LLMs isnt illegal. This is bad faith argument
You're talking about legality again. I'm talking about ethics.
Using LLMs for software development is a safety hazard. It also has a societal risk, because it centralizes more data, more power, more money to tech oligarchs.
It's ethical to fight this. Still not commenting on legality.
You're not forced to work there and use those tools. If you don't like it, then leave the job. Intentionally breaking things is unethical especially when you're receiving a paycheck to do the opposite.
Again, no one is forcing him to be there. He's breaking something on purpose. I think you should read up on ethics because this take "I don't like it therefore whatever I do is ethical" is juvenile.
That's quite the strawman. The reason it's ethical is not that LLM's are unpopular or someone dislikes them. It's ethical because LLMs introduce safety hazards, i.e. they cause harm.
That's extremely unethical. You're being paid to do something and you deliberately broke it which not only cost your employer additional time and money, but it also cost your customers time and money. If I were you, I'd probably just quit and find another profession.
That's not "sabotaged", that's sabotaged, if you intentionally introduced the bugs. Be very careful admitting something like that publicly unless you're absolutely completely sure nobody could map your HN username to your real identity.
They coasted on momentum for half a year. I don't even think it says anything negative about the current CTO, but more of what an exception JGC is relative to what is normal. A CTO leaving would never show up the next day in the stats, the position is strategic after all. But you'd expect to see the effect after a while, 6 months is longer than I would have expected, but short enough that cause and effect are undeniable.
Even so, it is a strong reminder not to rely on any one vendor for critical stuff, in case that wasn't clear enough yet.
You can coast for quite some time (5-10 years?) if you really lean into it (95% of the knowledge of maintaining and scaling the stack is there in the minds of hundreds of developers).
Seems like Matthew Prince didn't choose that route.
The problem is that CF operates in a highly dynamic environment and you can't really do that if the minds of those hundreds of developers relied for the major decision making on a key individual.
This is the key individual paradox: they can be a massive asset and make the impossible happen but if and when they leave you've got a real problem unless you can find another individual that is just as capable. Now, I do trust JGC to have created an organization that is as mature as it could be, but at the same time it is next to impossible to quantify your own effect on the whole because you lack objectivity and your underlings may not always tell you the cold hard truth for reasons all their own.
And in this case the problem is even larger: the experience collected by the previous guru does not transfer cleanly to the new one, simply because the new one lacks the experience of seeing the company go from being a tiny player to being a behemoth, and that's something you can do only once.
I've always been of the opinion that without JGC Cloudflare did not stand a chance, irrespective of those hundreds of developers. And that's before we get into things like goodwill.
And of those hundreds of developers you have to wonder how many see the writing on the wall and are thinking of jumping ship. The best ones always leave first.
I would not be surprised at all if this whole saga ends with Google, Microsoft or Amazon absorbing CF at a fraction of its current value.
been at cf for 7 yrs but thinking of gtfo soon. the ceo is a manchild, new cto is an idiot, rest of leadership was replaced by yes-men, and the push for AI-first is being a disaster. c levels pretend they care about reliability but pressure teams to constantly ship, cto vibe codes terraform changes without warning anyone, and it's overall a bigger and bigger mess
even the blog, that used to be a respected source of technical content, has morphed into a garbage fire of slop and vaporware announcements since jgc left.
Do you feel that Matthew Prince is still technically active/informed? I've interacted with him in the past and he seemed relatively technically grounded, but that doesn't seem as true these days.
Rather than be driven by something rational like building a great product or making lots of money he is apparently driven by a desperate fear of being a dinosaur.
Regardless of how competent he is or isn’t as a technologist, a leader leading with fear is a recipe for disaster.
I’ve had a lot of problems lately. Basic things are failing and it’s like product isn’t involved at all in the dash. What’s worse? The support.. the chat is the buggiest thing I’ve ever seen.
How about accurate billing info. The ux can’t even figure out we’re annually not monthly. Maybe the AI slop will continue to miscount resources and cost you revenue or piss off a customer when the dashboards they been using don’t match the invoice
You know what they say, shit rolls downhill. I don't personally know the CEO, but the feeling I have got from their public fits on social media doesn't instill confidence.
If I was a CF customer I would be migrating off now.
exactly. recently "if the cto is shipping more than you, you're doing something wrong"
cto can't even articulate a sentence without passing it through an LLM, and instead of doing his job he's posting the stupidest shit to his personal bootlicking chat channel. I cringe every time at the brown-nosers that inhabit that hovel.
no words for what the product org is becoming too. they should take their own advice a bit further and just replace all the leadership with an LLM, it would be cheaper and it's the same shit in practice
I have worked in some dysfunctional places but nothing like that, does sound bad.
Just got to keep your head, remember it’s just a job and you get paid regardless. Clock in, clock out, do the work assigned to you but mentally just check out while you look for a new role
I think both your statement and their statement are too strong. There is no reason to think LLMs can do everything a human can do, which seems to be your implication. On the other hand, the technology is still improving, so maybe it’ll get there.
2) There's no fundamental reason preventing some future technology to do everything humans can, and
3) LLMs are explicitly designed and trained to mimic human capabilities in fully general sense.
Point 2) is the "or else magic exists" bit; point 3) says you need a more specific reason to justify assertion that LLMs can't create new concepts/abstractions, given that they're trained in order to achieve just that.
Note: I read OP as saying they fundamentally can't and thus never will. If they meant just that the current breed can't, I'm not going to dispute it.
> 3) LLMs are explicitly designed and trained to mimic human capabilities in fully general sense.
This is wrong, LLM are trained to mimic human writing not to mimic human capabilities. Writing is just the end result not the inner workings of a human, most of what we do happens before we write it down.
You could argue you think that writing captures everything about humans, but that is another belief you have to add to your takes. So first that LLM are explicitly designed to mimic human writing, and then that human writing captures human capabilities in a fully general sense.
It's more than that. The overall goal function in LLM training is judging predicted text continuation by whether it looks ok to humans, in fully general sense of that statement. This naturally captures all human capabilities that are observable through textual (and now multimodal) communication, including creating new abstractions and concepts, as well as thinking, reasoning, even feeling.
Whether or not they're good at it or have anything comparable to our internal cognitive processes is a different, broader topic - but the goal function on the outside, applying tremendous optimization pressure to a big bag of floats, is both beautifully simple and unexpectedly powerful.
Humans are trained on the real world. With real world sensors and the ability to act on their world. A baby starts with training hearing, touching (lots of that), smelling, tasting, etc. Abstract stuff comes waaayyyyy later.
LLMs are trained on our intercepted communication - and even then only the formal part that uses words.
When a human forms sentences it is from a deep model of the real world. Okay, people are also capable of talking about things they don't actually know, they have only read about, in which case they have a superficial understanding and unwarranted confidence similar to AI...
All true, but note I didn't make any claims on internal mechanics of LLMs here - only on the observable, external ones, and the nature of the training process.
Do consider however that even the "formal part that uses words" of human communication, i.e. language, is strongly correlated with our experience of the real world. Things people write aren't arbitrary. Languages aren't arbitrary. The words we use, their structure, similarities across languages and topics, turns of phrases, the things we say and the things we don't say, even the greatest lies, they all carry information about the world we live in. It's not unreasonable to expect the training process as broad and intense as with LLMs to pick up on that.
I said nothing about internals earlier, but I'll say now: LLMs do actually form a "deep mofel of the real world", at least in terms of concepts and abstractions. That has already been empirically demonstrated ~2 years ago, there's e.g. research done by Anthropic where they literally find distinct concepts within the neural network, observe their relationships, and even suppress and amplify them on demand. So that ship has already sailed, it's surprising to see people still think LLMs don't do concepts or don't have internal world models.
Never ascribe to malice what can be sufficiently explained by incompetence. And i think it’s fair to say the best and brightest at Google aren’t turning their attention to YouTube lately. Except maybe to make training datasets for Gemini N+1 :)
The unspoken assertion that Rust and Python are interchangeable is pretty wild and needs significant defense, I think. I know a lot of scientists who would see their first borrow checker error and immediately move back to Python/C++/Matlab/Fortran/Julia and never consider rust again.
To be fair, triton is in active use, and this should be even more ergonomic for Python users than triton. I dont think it’s a sure thing, but I wouldn’t say it has zero chance either.
The idea that a GPU is dedicated to a single inference task is just generally incorrect. Inputs are batched, and it’s not a single GPU handling a single request, it’s a handful of GPUs in various parallelism schemes processing a batch of requests at once. There’s a latency vs throughput trade off that operators make. The larger that batch size the greater the latency, but it improves overall cluster throughput.
I work on a team that has actually deployed NN based surrogate models into production in industry. We don’t use PINNs for the simple reason that many industrial scale solvers are solving significantly more complex systems than a single global PDE (at least in CFD, perhaps other areas are simpler). For instance, close to the boundaries, the solver our engineers use uses an approximation that does not satisfy conservation of mass and momentum. So when we try to use physical constraints, our accuracy goes down. Even in the cases where we could technically use PINNs we find they are underwhelming, and spending time on crafting better training data sets has always been a better option for us.
reply