fancyfredbot

I bet there's a better way than unpacking, running sequentially and repacking. Even if the algorithm is very branchy you save a pack and unpack.

The relatively slow depreciation of GPU value is an artifact of supply constraints. If you run fp4 inference and could choose freely between Hopper and a Rubin, the performance per watt would make the Hopper unattractive even if you paid zero for the hardware and only for the power.

You can't get the Rubin, or even the Blackwell, so you will pay for the H100 but this won't last if fabs ramp up capacity.

s/banks/private credit/

Things have changed since 2009. It would be private credit which fails this time, not the banks.

Private credit is not supposed to be systemically important and it's not supposed to need bailing out. Maybe we'll find out how true that is in practice.

A lot of OpenAI/Anthropic compute actually belongs to the hyperscalers. They have long term contracts and will be able to charge some kind of margin for access to the compute they've secured. However I don't think this would last long as the hyperscalers will be able to undercut them.

The term vibe coding actually originated from a tweet from Andrej Karpathy and so there is actually a sort of definition of the term. He said vibe coding is where you "forget the code exists":

https://x.com/karpathy/status/1886192184808149383?lang=en

So if you "read the diff" it isn't vibe coding, at least not by Karpathy's definition.

Obviously it's not a "rule" and you can still call it vibe coding if you want to. Maybe the meaning of the term has evolved since his tweet anyway.

I increasingly feel that the reason vibe coding specifically prohibits reading the generated code is that it's impossible to forget the horror that lurks in these python files.

The code will do what you asked for in a broad sense but wherever choices arise on how to accomplish the goal it'll have made those choices incoherently.

There'll be strange validations applied inconsistently to some user inputs but not others. Data will be repeatedly sorted or converted to lower case for no reason at 3 stages. The column labels are all hard coded strings even though they are the first row of the input CSV. It'll create global functions taking data classes half the time and classes the other half.

If you look, and if you think you might one day need to update and maintain the code manually, then it's very hard to resist fixing this sort of thing, but fixing this sort of thing cancels a lot of the time saving out.

Buffoon (not bafoon).

They are taught the difference through reinforcement learning with verifiable rewards. Pretending you've solved the task or making up a story about how you solved it won't do well in that training step.

Normally people refer to the compute-bound phase as "prefill". Nothing wrong with saying it's building the kv cache though, it's accurate just unusual.

Having Mojo support multiple platforms creates incentive to adopt Mojo and therefore write code in a language which can compile and run on Qualcomm hardware. This is good for Qualcomm.

However the danger is that the language sees wide adoption but nobody uses it with Qualcomm hardware. Instead it might encourage people to buy AMD. This is a terrible outcome for Qualcomm. They paid to boost someone else's sales.

So the incentive is to make sure it runs best on Qualcomm and to at least slightly hobble other hardware. But the safest thing overall is to support Nvidia, Qualcomm, and that's it.

The top end of the range they've shown here for A2W will assume you only heat the water to a very low temperature, and have a huge underfloor radiator. If you do that, and you only heat the water to 35C, A2W can be more efficient.

Most UK houses have radiators and not underfloor heating and so cannot operate in that way without a hugely expensive retrofit of underfloor heating. It's typical to heat the water to over 50C if you want to reuse installed radiators. In this configuration the A2W pump is less efficient than A2A.

Worth noting these numbers are low. Latest A2A pumps can have COP of up to 4, but an A2W can achieve 4.5 if you keep water temps very low.

Modular now joins SYCL, OpenCL, and One API on the list of cross platform languages which never really became cross platform.

After so long and so much investment in AI, the best cross-platorm API we've got for high performance Kernels is vulcan, a graphics API. That is sad.

Still, this is pretty good for Modular's employees, probably good for Qualcomm. It's just terribly disappointing for anyone who invested time learning mojo in the hope it might actually become cross platform.

Air to water heat pumps are typically less efficient than air to air. Both will work in cold UK temperatures. However the water is often heated to a higher temperature which can reduce efficiency.

Question one: How much did this cost OpenAI?

Question two: Why are OpenAI spending that money taking talent from Google, who can definitely outspend them for talent, and not Anthropic, who are leading the market and are at least somewhat financially constrained.

I don't think my logic would lead to only one country with a launch rocket. Russia and China do not really have the option of using American rockets.

Britain however does have this option and as far as I know it has never had a problem launching satellites, so I guess it's worked out so far.

The way I thought it worked was that congress would set a budget and scientists would decide how to spend it.

Perhaps naively I thought these scientists would want to do science and would be unwilling to steer funds away from whichever projects they liked in order to fund the removal of some sensors.

I guess maybe the scientists who make these decisions are also partisan and happy to do as the administration asks.

I understand the present US administration would want to stop funding this, and that they have the power to do so.

I don't understand how that has led to the sensor network being dismantled. Surely it would have been cheaper to leave it in place and stop maintaining it?

I was suggesting US models are very good, and would probably be preferable to a less capable sovereign AI. My point was that this makes building European sovereign AI unappealing.

Similarly I was not suggesting Chinese models are an alternative to sovereign AI. An alternative which might be able to fulfil some of the objectives of a real sovereign AI while being far cheaper.

Forget regulations, would it make sense for Europe to train a frontier model of it's own? Would it be sufficiently better than fine tuning a Chinese model? Would it actually be competitive with US frontier models? Would enough people pay to use it even within Europe to pay for the training costs? Do we have enough inference capacity that enough people /could/ use it? Would being "European" allow any governments in Europe to trust it, rather than deciding that actually there needs to be a French, German, Italian, Spanish and UK sovereign AI?

I am guessing that enough of these questions can be answered with "no" that nobody really wants to invest.

For the same reason there isn't really a serious third start up competitor to OpenAI and Anthropic.

Britain didn't abandon it's space programme. It abandoned a launch rocket programme though. That was over 50 years ago and the rocket was less capable and more expensive than alternatives at the time.

The article is right that open models already compete well with the frontier labs, and that the main thing holding big corps back from switching is fear of China.

I can't see OpenAI or Anthropic undermining their business by releasing top tier open models, but surely Nvidia will do it eventually.

In other news, the gold rush has entered a new phase as miners pivot to selling shovels.

For now. SpaceX will be acquiring Tesla as soon as Elon gets around to it.

It's circular when money flows from A to B and then back from B to A again.

This is a series of transactions in which money flows from Google to SpaceX. There is no flow of money from SpaceX to Google. So it's not a circle.

I saw one useful feature added, which was support for lower resistance tips. You can also get it to report on the input voltage which is marginally useful if running on battery.

Most of it seems to be people having fun. Animations, enabling Bluetooth, web interfaces over aforementioned Bluetooth, etc. That's all pretty cool and hacker news is hardly the place to question why you'd want to do it, but I'll admit I find the number of people interested in hacking on a soldering iron a bit of a surprise!

Most large companies have the CEO answerable to a board elected by shareholders. CEO still has a lot of power but there are some checks and balances.

Zuck and Musk are somewhat exceptional in being dictator-CEOs.

Good news - Anthropic will pay you £600k to build the AI which replaces you.

Drag it out for a couple of years and you'll be set.

If you pay for any tier of copilot then you will get unlimited access to the auto complete.

If you don't use agents it's possible you'll not notice much of a change to copilot

Capital is cheap for Google because they are making a lot of money and they have very little debt.

They are considered a very low risk and can borrow for a long time at low rates. They recently issued a 100 year bond.

They seem to have decided to issue equity rather than borrow more. This is probably so that they can maintain the ability to borrow very cheaply in future if necessary.