You probably have other stuff on your mind right now, thus I can understand if you are not in the mood for answering, but I‘m too curious to not ask:
According to the Reuters article, AI use has increased 6x over only three months. How did that feel from the inside? I’m especially curious because Cloudflare is not a toy company, and this is not about some influencer trying to sell me their latest „this changes everything“ bullshit.
So, shifting a company significantly towards agentic AI, and I assume this isn’t simply about „install Claude Code on every desk“: would you say it actually works? Or would you say it’s still more of a bet, and still needs to prove itself as a sustainable long-term strategy?
I work at a similar scale company. Like an average person's experience, some things are amazing and super productive with AI and some things aren't. And it's not always the same things all the time.
Sometimes we are able to do a ground up rewrite of a service and squeeze huge efficiency gains out of it all bc AI is helpful in doing so and we have a very good test harness.
Sometimes it makes subtly wrong suggestions that people follow and cause outages.
Sometimes it leads to huge headaches for devs who have to review huge backlogs of code with no idea which parts are serious and which are low effort AI slop.
Sometimes it lets you do a 2 month project in 2 weeks.
This is only a tangential question, but anyway: I‘ve read several years ago that since around Photoshop version 4, 99% of the work is about keeping the application UI usable with all these new features, and not about „hard“ technical challenges within the features themselves. Is that true?
That sounds plausible. Most of the features are kind of gimmicky bolt-ons added piecemeal and not really integrated with each-other. They make for cool 10-second demos but then most users ignore them because they aren't part of a coherent system. The result is a menu after menu of gimmicks, like a cabinet of hyper-specialized kitchen tools bought from infomercials. There has been limited product vision about the core abstractions and their basic composability. If you give a skilled user a photoshop version from the early 2000s they'll largely be able to do what they need, because there hasn't really been much fundamental innovative improvement in the past ~25 years.
Microsoft actually does a fairly good job with this. Here's a part of a talk that goes over a single feature in PowerPoint (a slide animation that morphs the contents of one slide into a different slide) and demonstrates how this feature interacts with the enormous existing PowerPoint feature set in interesting ways. https://youtu.be/_3loq22TxSc?t=1409 It's obviously a stupid gimmicky feature but whatever team Microsoft put on it were clearly overachievers.
Thanks for sharing that fascinating video! It seems like a fair bit of work went into it. One criticism I have is that it is undiscoverable and opaque; it is not obvious how it is going to behave. I wonder how many users are aware of it.
There is a lot of effort put in to making the application usable, no question. At the same time, we have added a litany of new features and tools since Photoshop 4, many of which I would describe as extremely technically challenging.
Reading only the parts of the post that are not about AI does not instill the sense that Mr Armstrong is the kind of person who would hesitate to say that people are let go because the company wants/needs to save money.
What was eye-opening, or rather, sobering for me was when I read an interview with an engineer who explained how incredible difficult it is for a robot to orient itself when it is lying on the floor and wants to stand up.
Yes, it can do the required motions just fine, that’s not the point. But think about yourself when you are lying on the floor: it’s really easy to determine if this is safe, if you are lying underneath something and so on. You just feel that.
A robot cannot do that; all they can do is look around as good as possible and visually determine their situation.
It is also things like "I can feel that my left knee is bearing a little too much weight, I should shift weight to my right hand and use that to push myself up" - things that come automatically to animals after learning the hard way in infancy (some of it is innate; baby animals are clumsy, but usually more mobile than human infants). Regardless of learned-vs-instinct, these abilities rely on sophisticated "sensors" and cognition. I suspect engineering the sensors is actually a bit harder, but I'm also not optimistic about a deep learning approach to the cognition.
A significant underappreciated advantage of animals over AI: lifeforms can "learn the hard way" more easily than 2020s robots because of cheap self-repair. AI labs are reluctant to damage their robots, but an essential part of humans learning to move safely is severely bonking your head and reckoning with the consequences - "hey, dummy, why did you trip and fall and bonk your head? Because you were running like an idiot."
I am learning the hard way to this day :) I have been practicing with work knives. A few months ago I got stupid and impatient, and sliced my thumb nastily. If I didn't block the cut with my thumbnail (still ruined) I might have chopped bone. It is hard to say precisely what I learned from this experience - "don't be stupid and impatient" is facile - but I know I learned a lot. I am actually optimistic about targeted surgical robotics. But for a general-use humanoid robot, I would not want to give it a knife if it's not capable of feeling pain. I never use big knives anywhere near my cats because I understand intuitively that they are nimble and unpredictable and easily stabbed by knives. I didn't need to be trained on this. A robot kind of does. Yikes.
Yeah but imagine yourself lying on the floor with your vision being your only sense, plus an info floating in your mind: „fyi, you are no longer upright“.
That’s all, you feel nothing else. Now your job is to move all parts of your body in just the right way.
The point about being aware of lying underneath some object was interesting. Sound might matter, like the frequency of background noise changes when you're in an enclosed space, and listening to your own shuffling noises helps you know when you've planted your feet right - or something. I have some really effective ear plugs and I notice they make it harder to move around.
Having said that, I've probably hit my head on the underside of an open cupboard door five or six times in my life, and I expect to do it again.
Would you mind sharing what you can and want about how the sausage is made? I would love to hear concrete cases where actual leverage is measurable. I‘m asking in good faith, not to attack your standpoint.
I‘ve built https://github.com/dx-tooling/productbuilding-orchestration for more or less the same use case as Broccoli, but in my case a) the „frontend“ is Slack, not a ticket system (but the agent people talk to in Slack manages the ticket system under the hood), and b) everytime there is a new commit on a PR, a preview system is bootstrapped or updated, and the codebase that is to be previewed explains how it wants to be deployed for preview.
I think it totally depends on the team on what the correct workflow is. The agentic workflow should accelerate what you have instead of asking your team to adopt something new. This is why we think it helps to own the orchestration yourself so you have that level of control!
(I‘m saying this as someone who uses AI for coding a lot and mostly love it) Yeah, but is that really the same? Compilers work deterministically — if it works once, it will work always. LLMs are a different story for now.
Said another way, compilers are a translation of existing formal code. Compilers don't add features, they don't create algorithms (unrolling, etc., notwithstanding), they are another expression of the same encoded solution.
LLMs are just translating text into output, too, and are running on deterministic computers like every other bit of code we run. They aren't magic.
It is just the scope that makes it appear non-deterministic to a human looking at it, and it is large enough to be impossible for a human to follow the entire deterministic chain, but that doesn't mean it isn't in the end a function that translates input data into output data in a deterministic way.
LLMs are deterministic, too. I know there is randomness in the choosing tokens, but that randomness is derived from a random seed that can be repeated.
Only if the seed is known. Determinism is often predicated on perfect information. Many programs do not have that. Their operations cannot be reproduced practically. The difference between saying deterministic and non-deterministic is contextual based on if you are concerned with theory or practicality.
llama.cpp is deterministic when run with a specified PRNG seed, at least when running on CPU without caching. This is true regardless of temperature. But when people say "non-deterministic", they really mean something closer to "chaotic", i.e. the output can vary greatly with small changes to input, and there is no reliable way to predict when this will happen without running the full calculation. This is very different behavior from traditional compilers.
No, LLMs ARE deterministic, just like all computer programs are.
I get why that is in practice different then the manner in which compilers are deterministic, but my point is the difference isnt because of determinism.
Fair, although you can absolutely use local LLMs in a deterministic way (by using fixed seeds for the random number generation), and my point is that even if you did that with your LLM, it wouldn't change the feeling someone has about not being able to reason out what was happening.
In other words, it isn't the random number part of LLMs that make them seem like a black box and unpredictable, but rather the complexity of the underlying model. Even if you ran it in a deterministic way, I don't think people would suddenly feel more confident about the outputted code.
LLMs are deterministic[1], but the only way to determine the output is to empirically run them. With compilers, both the implementor and a power user understand the specific code transformations they are capable of, so you can predict their output with good accuracy. I.e. LLMs are probably chaotic systems.
edit: there might be a future where we develop robopsychology enough to understand LLM more than black boxes, we we are not there yet.
[1] Aside from injected randomness and parallel scheduling artifacts.
Global copper reserves are about 980 million tonnes, with 1.5 billion tonnes of identified resources and 2024 mine production of about 23 million tonnes.
A conventional car uses about 23 kg of copper and a battery-electric car about 83 kg, a difference of roughly 60 kg per vehicle.
With more than 17 million EVs sold in 2024, that implies about 1.4 million tonnes of copper embodied in those vehicles, or about 1.0 million tonnes more copper than comparable conventional cars would have used; applying the same assumptions to the current global passenger-car fleet implies roughly 120 million tonnes total or 87 million tonnes incremental copper for an all-BEV fleet.
Separately, the IEA says that under today’s policy settings and announced projects, copper faces an implied 30% mined-supply shortfall in 2035, while expanded recycling could reduce new mine needs for copper by about 35% by 2050.
Peak oil production was in 1970 so I don't know what to tell you other than it's basic physics. You should look into something called EROEI (energy return on energy invested). You might find it useful for your next argument about the viability of EVs.
I don't know what to tell you. ICE cars are doomed, it's basic physics. They depend on something that is a finite resource that you agree peaked a long while ago. They're well along their way to complete failure. And yet while we'll still have the copper to recycle into new motors we won't have more oil. But here you are telling me we're going to run out of copper despite lots of sources telling you that's absolutely nonsense, but you'll believe it because some guy in a YouTube video told you "that's just basic physics, there's only so much copper, gotcha!"
> You should look into something called EROEI (energy return on energy invested).
I'm very familiar with it. As easy supplies of oil are getting used up the EROEI of oil is falling. It takes a lot of energy to transport and refine all those tar sands after all. It takes a lot of energy to cryogenically transport LNG around the world. And once you burn it, you need yet another shipment of it around the world. Meanwhile the EROEI of modern renewables is often even exceeding natural gas plants. I don't know what to tell you other than it's basic physics.
You might find actually reading sources other than the oil and gas industry lobbyists enlightening on the topic. Quit asking the Altria group if you should quit that smoking habit and go talk to an actual doctor. And before you brush that off, it's exactly what you're doing. You're asking an oil industry insider if you should bother looking into buying less oil. Of course he's going to tell you nothing else makes sense, keep buying oil.
You should look at actual facts with real figures for your next argument about EVs instead of claiming there's not enough copper. I'm sorry you've swallowed so many obvious lies by these guys.
That's right, once you've realized an EV charged from solar power will have a higher EROEI than a gas car fueled from oil sands and corn juice "the discussion has run its course", and that gasoline has a finite time of economic usefulness rapidly approaching "this discussion has run its course".
I dunno what to tell you, it's just basic physics. Unless you're paid millions from the oil industry, in which case it's suddenly very complicated.
I say this as someone who owns two EV cars, zero ICE cars, and loves everything about owning and driving EVs: it baffles me how quickly and noticeably consumers shift their preferences; I think I read something along the lines of „for the average German commuter, the petrol price spike means 6 euros plus per week in spending“ — and that is enough for so many people to go „okay screw it I’m switching to new technology for my planned purchase“?
I mean, great that it happens, but yeah, I‘m baffled.
It’s the stress around it, not the current price. That’s what people are replacing.
If I’m trying to plan for the future in a world where conflicts this destructive are permanently on the menu, I’m not going to ever buy an ICE car again. No one wants to be at the mercy of anyone else where possible and as someone who only owns ICE cars, it’s been very stressful few weeks.
I think you are on to something here. But it would mean that sentiment could change on a whim: a serious blackout somewhere in Europe that makes continent-wide headlines, and EV demand might crash.
Given the lack of blackouts around here, I've not bothered to investigate if home PV systems can still charge batteries, or if the inverter dependence on the grid for phase-locking the AC prevents that.
There's already been news stories about people powering their homes from their EV during blackouts, the oldest example I know of this from 2012 albeit with a hybrid and it was burning petrol: https://cleantechnica.com/2012/11/10/man-lights-house-with-t...
> Given the lack of blackouts around here, I've not bothered to investigate if home PV systems can still charge batteries, or if the inverter dependence on the grid for phase-locking the AC prevents that.
The technology is called "islanding mode". It adds a little extra cost to the inverters, and you have to specifically request it, but the tech exists.
The second part is the fact that domestic brands all came out with good EVs and many people are installing home solar which makes savings even more drastic.
According to the Reuters article, AI use has increased 6x over only three months. How did that feel from the inside? I’m especially curious because Cloudflare is not a toy company, and this is not about some influencer trying to sell me their latest „this changes everything“ bullshit.
So, shifting a company significantly towards agentic AI, and I assume this isn’t simply about „install Claude Code on every desk“: would you say it actually works? Or would you say it’s still more of a bet, and still needs to prove itself as a sustainable long-term strategy?
reply