SilverBirch

It's computer Gain of Function research.

The "move to the us" thing seems to just be absolute twitter brain. The biggest problem London founders face is raising funding and these guys seemingly raised plenty based on an idea that that didn't seem to understand the market they were targeting.

I think this is such a great example of misunderstanding a political problem as a technical problem, and misunderstanding twitter as real life.

the political winds seemed favourable

Sorry. How? Some libertarians on twitter were in favour. But if you have spent any time paying attention to British politics at all you would note the incredible sway the elderly have over politics. Even the people who are pro-building aren't stupid enough to think you're going to achieve that by building on green belt land, so this idea of an uplift of 140x is just farcical. Half of British debate on house building is disingenuously accusing your opposition of concreting over the green belt, whilst disingenuously promising you can build millions of new homes without concreting over the greenbelt.

The problem with building in the UK is mainly planning permission. In the UK is you need approval from the local council and the local council is elected by Nimby pensioners. So you can't get approval. That's it. That's the problem. That's a political problem that you can solve by getting central government to remove local governments ability to block projects. Or by running targeted political campaigns in local areas to get specific things approved. Not by automating the form your fill in to get your planning application rejected.

Isn't this the opposite of what I want? I don't want people willing to pay getting into my inbox. Those are the people who think they can get more from me somehow. Those are exactly the people I want to not be in my inbox.

Don't copy that floppy! - https://www.youtube.com/watch?v=up863eQKGUI

If Rivian’s native UI is so great, then their customers… won’t use CarPlay. It’s that simple.

I kind of disagree with this. Airpods are purely additive, customers can just choose to use different headphones with their iPhone if they want. But they don't want, because Apple lets Airpods interact with the iPhone in a way that other manufacturers can't.

So no, carplay wouldn't be mandatory but it's likely that Apple's leverage will kill their in house offering.

To be honest I don't think what the models themselves say in relation to these specific questions matter. Because I don't think it reflects are durable underlying worldview. I suspect that the way you frame things is going to influence them so muc that it's irrelevant what they would say when put in a petri dish.

What is a lot more important is how they're develop. To take the two sides of the spectrum - they say has a slightly expansive attitude towards civil liberties, but if you try to use it's tool it will phone it's owners and ask permission for you to use it. Or you can pick up Grok one day and find out that Elon Musk had a bad weekend and Grok is back to being mecha hitler.

Google acquired his company in 2024 for $2.7Bn with him taking about 40% of that. I'm quite sure that no matter where he went, any lab or his own start up, he would be fine financially.

I totally understand the concern that forcing a digital proof of age turns into you having to go to some 3rd party who won't actually just provide a "Yes, this person is 18" style verification but instead will turn into user tracking and ad networking etc. But is that actually required by the bill?

Because what I'm getting at is Apple is already privacy focused. It would be entirely plausible to me that we end up in a situation where Apple's implementation of this is absolutely just "Yup, here's the token that proves I know this is an adult" and if doing that screws Meta's plans for advert-nirvana, I would expect them to take that route.

With the porn ban in the UK we have this worst of both worlds at the moment. We have forced ID verification by these creepy third parties, but it's so fragmented and localized that no big player like Apple is stepping in to provide anything more consumer friendly.

I feel like a lot of this advice is kind of dangerous. How do I draft a tight investor memo? I'll ask the slop machine!

It's kind of analogous to how I'm writing code right now. For simple stuff or low priority stuff I'll fire claude at it and won't look at the code if it works. But for the important stuff I'm very carefully integrated into the cycle making sure what's coming out at the end is just right. I'm carefully constructing prompt loops and validation cycles to make sure what comes out looks like what I want - because I have the knowledge and experience of what works for my specific use case. Drafting an investor memo seems like the second category of thing, you need it to be right. I don't think claude offers much of value there. What's more - if you start slopping your investors, you are going to piss them off. Unless Claude is going to say it has some special data source it's used to train on so it knows good from bad, I think this is a bad idea.

This article also kind of fits in the category of "Here's how to use AI for EVERYTHING!" and actually it would be far more valuable to say "This is the bits that AI is good at, and here's where you need to do it yourself" - which is obviously a position that Anthropic can't hold.

Just to be clear, this is the same reason that social media companies don't tell you about how they detect spam and they create shadow bans and things like this so that people don't know they've been detected and figure out the mechanims.

And it doesn't work. Even a bit. It's a constant constant cat and mouse game. Maybe they can slow people down slightly, but they won't be able to stop them, and good luck protecting yourself from Elon Musk snooping your stuff in his data centre.

What are you even classifying as accurate or correct? Do you take every 51% prediction from FiveThirtyEight and if the result is a win you consider that forecast accurate? And every 49% prediction must result in a loss? This just not how statistical forecasts work.

What it would be reasonable to say is if his model had correctly predicted the outcome of a significant sample of elections, then you could say his model has some accuracy or predictive power.

I don't know why you're couching that in a hypothetical, FiveThirtyEight has repeatedly done that exercise.

But it still would never have been accurate or right in the specific instances it got wrong

It is core to the concept of a probability that the result is going to go the opposite way from the prediction sometimes! It's meaningless to call it "wrong".

To give you a trivial example: The simplest way I can put this is that turn out varies based on the weather[1], and turn out is skewed by party. So if it rains on election day you are going to get a different result, and that result can flip the outcome of the election if the election is close. So it’s kind of a nonsense to say. “Trump would have won 100 times out of 100”. Are you saying Nate Silvers model should have had a perfect meteorological model to predict the weather? Or are you saying the election wasn’t close? In which case you’re just wrong on the facts.

The 70% figure is saying “we know most of the information needed to determine what the outcome of the election will be but we don’t know everything so can’t be certain”. There is no process where you can know every factor that determines the result in advance with absolutely accuracy and I don’t know why people expect there would be.

[1] https://www.sciencedirect.com/science/article/pii/S026137942...

I think it's unquestionably right that these companies can't all win, and those that don't win are going to burn a lot of money for nothing. However there's kind of two directions this can go: Compute gets cheaper, in which case there's no monopoly it'll be easy for many companies to make good models and there won't be pricing power on serving a good model. The other case is compute gets cheaper but we keep using more and more of it, so it does likely become winner take all. The first scenario is good for the economy but likely bad for the returns on these AI stocks. The second is maybe bad for the economy and maybe not even good for the winner.

Take Google or Meta: Today Google makes a shit-tonne of money and to make that money they need to run some servers. The servers are extremely cheap relatively to the revenue they make running the business. This makes them a very attractive stock - the core of why SAAS looks great. Now let's assume the monopoly path. Google can win. I think they likely will win. But now they're going to spending... how many hundreds of billions constantly training new models? The cost of providing the service suddenly isn't small relative revenue they're getting. So even for them it looks awful for their valuation.

One aspect of my job is that I have a lot of autonomy and the work I do is such that I could push something out to the production environment and cause massive problems. We have processes in place to make sure that doesn't happen, but they're not robust processes, if you really wanted to you could get something out there that is harmful to the company. Now, there are two ways of looking at that - one is that it's really important to have robust processes to make sure that doesn't happen. But the other is you need people who understand that responsibility and take it seriously and whose personal values are such that they aren't just going to carelessly do stuff. At the end of the day the processes are only good if they're followed.

So one of the things I strongly look for when hiring is for people who have a high sense of personal responsibility. They're not going to just throw shit out there because it's easy or quick. They know they are responsible for what goes out and they really are going to own that responsbility.

In the same way, take a look at anything senior management says about their ICE or military contracts. It's not that I think they're doing something bad or that the military shouldn't have access to good technology. It's that at best they seem entirely disinterested in that what they're doing could be harmful or that they have any responsibility if it is.

It's not that I think Palantir is helping the US government bomb Iranian school chilren. It's that I don't think it would bother them if they were.

You're forgetting that xAI and X.com have both already been folded into SpaceX (First xAI acquired X.com, then xAI got acquired by SpaceX, both mergers were all-stock acquisitions so they were done with funny money). So when people say "SpaceX" now that does encompass both xAI and X.com as well. The reason Tesla wouldn't do this is because Tesla is a public company so it's more difficult for them to do insane shit without being sued.

Just commenting here to impact the controversy score.

Frankly I think it’s kind of childish to just put up a massive Uk wide block on your website. “Call your representatives”, ok dude, can I give you a list of things I want to change about your country’s policies?

Needs a [2010] tag. In almost all modern hardware development you'll have coding guidelines along the lines of "Always use blocking assignments for comb logic, always use non-blocking for sequential logic". You end up back at the same place as VHDL, by nature SystemVerilog is much weaker typed than VHDL. So you have to just have conventions in order to regain some level of safety.

What do you mean by simulate? Do you want the language to be aware of the temperature of the silicon? Because I can build you circuits whose behaviour changes due to variation in the temperature of the silicon. Essentially all these languages are not timing aware. So you design your circuit with combinatorial logic and a clock, and then hope (pray) that your compiler makes it meet timing.

The fundamental problem is that we're trying to create a simulation model of real hardware that is (a) realistic enough to tell us something reasonable about how to expect the hardware to behave and (b) computationally efficient enough to tell us about a in a reasonable period of time.

I guess this really depends on your view of the world. Was Marc Andreessen some visionary without whom no one would've ever figured out images could appear on websites. Some kind of Albert Einstein of cat gifs. Or was the img tag an inevitability once the web had enough bandwidth to transfer images.

There are obviously tonnes of accurate stereotypes in the TV show Silicon Valley, but one of the ones I think about often is when Richard calculates how much money Russ Hanneman has made investing his billions... and it works out to less than sticking it in the bank.

You've got all these silicon valley guys running around "venture investing", the truth is it's more of a life style than a money making exercise. They made their money decades ago, and now they're just sort of hanging around desperately trying to tell everyone how clever they are.

I'm always a little surprised at how these Tech CEOs are willing to go on TV and just spout nonsense. Firstly, 40% of college educated white women voted for Trump at the last election. Secondly, isn't the entire theory of Trump's support amongst working class voters an appeal to economic populism due to an erosion of their economic position? Aren't you literally describing a process that last time lead to a massive political shift in favour of those who were negatively economically impacted? Oh and you think all the white collar workers are going to lose their jobs, but you don't think that's just directly going to cause a recession that wipes out blue collar republican jobs?

It's difficult to (a) see how he can say this having given any real thought at all and (b) understand why he's going to on news interviews and winging it.

Whilst this is interesting I find the topic bought up on odd lots is more interesting. The idea was this: Once you've built a model, if you can sell tokens for a profit, this is a great business - just sell more tokens. But you can't just build a model and sell tokens. You need to build the best model to sell new tokens. So the question is much more "How much does it cost you to build a new SotA model" and then "How effectively can you monetize it". And since you need a SotA model, your only option if you have a bad model that isn't selling is to invest billions more into building a better model whose tokens you can sell.

So this turns into a death march.

If you are behind, the only thing you can do is make massive capital investments to catch up. Once you're ahead you can sell tokens until someone else catches up. And, breaking the model of normal of places like chip fabrication, your billions of investment may only keep you ahead for 2 months. So you have a tiny window to sell those tokens.

What you are talking about isn't inference cost. Yes, fundamentally what matters is all the work that goes into the models, including R&D, training, and inference.

But we talk about inference separately for a reason: largely inference cost is the scaling cost. Once you have a model the margin on your inference is how you get to profitability, as long as your margin is positive you can make the entire enterprise profitable by just selling more tokens. This is the same fundamental business that chip fabs work on. Yes it costs them a lot to get to the next node, but what's important is the margin they can get on the wafers they sell, because they sell tonnes of wafers.

It's pretty core to the concept of SAAS businesses that yes, you do consider all costs. But you want to focus on the margin of the bit that scales. This is why WeWork exploded, the thing they were scaling only scaled up at negative margin.

The point is that if their inference margin is positive, they can "just" scale up and become profitable. If their inference margin is negative, then scaling up the business actually causes problems.

There's two points here. The first is that a strategy of monetizing models to fund the goal of reaching AI is indistinguishable from just running a business selling LLM model access, you don't actually need to be trying to reach AGI you can just run an LLM company and that is probably what these companies are largely doing. The AGI talk is just a recruiting/marketing strategy.

Secondly, it's not clear that the current LLMs are a run up to AGI. That's what LeCun is betting - that the LLM labs are chasing a local maxima.

I think the big take away here isn't about misalignment or jail breaking. The entire way this bot behaved is consistent with it just being run by some asshole from Twitter. And we need to understand it doesn't matter how careful you think you need to be with AI, because some asshole from Twitter doesn't care, and they'll do literally whatever comes into their mind. And it'll go wrong. And they won't apologize. They won't try to fix it, they'll go and do it again.

Can AI be misused? No. It will be misused. There is no possibility of anything else, we have an online culture, centered on places like Twitter where they have embraced being the absolute worst person possible, and they are being handed tools like this like handing a hand gun to a chimpanzee.

I think you're missing the point. That phrase isn't giving a direct instruction to the chatbot to make sure it doesn't get elected to congress and subsequently pass laws prohibiting speech. That phrase is meant to tell it "You should behave like those guys on twitter who really want to say the N word, but have no problem with Kash Patel bullying Jimmy Kimmel off the air.

The data in the chatbots dataset about that phrase tell it a lot about how it should behave, and that data includes stuff like Elon Musk going around calling people paedophiles and deleting the accounts of people tracking his private jet.

Ah I see, so the misaligned agent was unsurprisingly directed by a misaligned human. Good grief, the guy doesn't seem to realise that starting your soul.md by telling your AI bot that it's a very important God might be a bad idea.

"Social experiment" you might as well run around shouting "is jus a prank bro!".

That can't/won't happen. Musk's wealth is primarily in SpaceX now and he has a much higher ownership stake in SpaceX than Tesla. As well as that, Tesla is public so he can't just do napkin math and decide to merge them. So the question is: Does Tesla buy SpaceX? Well no, Tesla can't afford it. Ok, well can SpaceX buy Tesla? Well no, SpaceX can't afford it either. So do they announce a merger? Well that doesn't make any sense because Tesla is valued like a meme stock so it would massively dilute Musk's ownership of the overall company. So the idea that they fuse might be driving up the stock, but by driving up the stock you're actually preventing it happening. If Tesla starts to trade at realistic multiples and comes down to lets say a 200Bn company, I'd expect SpaceX to snap it up at that valuation, but it'd be crazy to do it before then.