martinald

Yes fair enough, but try feeding my comment in :). It should be enough for it to go on. Then ask it to explain the concepts I mentioned and ask it to suggest follow-up questions for you to learn more about llama.cpp/local inference!

I've had best results with opencode. Running locally w/ 64GB RAM and Radeon 9070XT (16GB). NVidia should be easier (CUDA), I'm on Linux full time now but used to use WSL2 all the time and had all this working in it.

Yes I think you're right. Reinforcement learning is extremely compute heavy, which cursor doesn't have. And X.ai doesn't have the coding agent data anthropic/OpenAI has, but does have the compute.

However, one thing in AI is that while the usage goes up extremely quickly, it tends to go down just as fast. I know a lot of companies that are in the process of switching from Cursor to Claude Code, so in 6-12 months I'm not entirely sure of the data quality/quantity.

Also I think it is telling that they are calling them SpaceX not X. The X brand is absolutely toxic, especially in enterprise.

check out https://www.leaseloco.com/ (there's other sites - I am assuming you are in the UK), they have a lot of deals on. If you want a really good deal then I don't think you want to look for a specific car, you want instead to see what deals are going on and choose between them. It _seems_ that from time to time the manufacturers do a bit of a firesale via leases.

As sibling comment posted, if you run a company you can write off the cost of the lease against corporation tax (25% tax saving), VAT (a further 10% saving if you have personal use, otherwise the full 20% if it's purely business), and you don't have to pay income tax on the money (albeit with a small benefit in kind payment).

If you're in the higher tax brackets this means a £200/month lease (say) ends up more like £90/month.

And because of the "new" £3kish subsidy the govt put in, the car finance companies seem to basically apply that as a big discount to the lease value (or it seems that way?). So you could get a brand new ~£30k EV with no upfront payment and a 2-3 year term for <£100/month including maintenance. Mine even came with a free car charger install at home.

Not quite. The UK govt has a rolling target of how many EVs are sold as a proportion of all vehicles. The manufacturers get "fined" £15k per ICE car sold over the target. The targets are ramping up rapidly - 22% in 2024, 28% in 2025, 33% in 2026, etc, reaching 80% (AFIAK, keeps changing) in 2030. There's various trading mechanisms in place and what not so it's a bit more complicated than this.

So it's far better to sell EV below cost (Chinese or not) to get more sold than have to a pay £15k for an ICE car.

The Chinese manufacturers are arguably at double advantage here as they have more BEV to sell so it's far easier for them meet the targets, and they can 'sell' the excess to the Western manufacturers (and further subsidise their EVs!).

I'm not personally against it, I got a brand new EV on a lease recently for close to free after all the tax advantages, and it's not like the Western manufacturers didn't have time to prepare...?

Try using a MoE model (like Gemma 4 26b-a4b or qwen3.6 35b-a3b) and offload the inference to CPU. If you have enough system RAM (32GB is a bit tight tbh depending on other apps) then this works really well. You may be able to offload some layers to GPU as well though I've had issues with this in MoE models and llama.cpp.

You can keep the KV cache on GPU which means it's pretty damn fast and you should be able to hold a reasonable context window size (on your GPU).

I've had really impressive results locally with this.

I'd strongly recommend cloning llama.cpp locally btw (in wsl2) and asking a frontier model in eg Claude code to set it up for you and tweak it. In my experience the apps that sit on top of llama.cpp don't expose all the options and flags and one wrong flag can mean terrible performance (eg context windows not being cached). If you compile it from source with a coding agent it can look up the actual code when things go wrong.

You should be able to get at least 20-40tok/s on that machine on Gemma 4 which is very usable, probabaly faster on qwen3.6 since it's only 3b active params.

I suspect this is the best option. Focus on one city pair to start with, _you_ are the courier and find customers on that city pair. Then you can start figuring out how to attract people onto that city pair so you don't have to do it anymore (as there will be demand).

Atlassian just goes from misstep to misstep. I still use their products quite often. The amount of P0 bugs I experience is absolutely crazy:

- Bitbucket workers are hopelessly out of date (self hosted). We've had to put so many random workarounds in especially for Docker, as they don't keep them up to date enough

- I have had a bug in JIRA for years where I can't reorder a new ticket unless I refresh the page

- Every new feature they introduce into JIRA/Bitbucket over the past couple of years just doesn't work.

- I tried their AI stuff on the free trial, didn't work at all, tried to cancel, can't cancel the free trial online and had to write a load of support tickets (of which the support ticket contact form bugged out multiple times).

Anyone have any insight into why things have got so so dysfunctional? Tech debt? Talent leaving? Both? Even 'bad' enterprise software tends to be able to keep the most basic features running, but Atlassian is a whole new category. If you check their 'community' it is just hundreds/thousands of bugs with workarounds.

Yeah, though AFIAK these event loops still suffer from blocking on (eg) complex json parsing or anything CPU driven (where real multithreading shines).

But regardless I agree, I'm just saying that these kind of patterns _are_ needed in any moderately complex system, and taking the view that "it's great not to even have it" in the core framework is really strange to me. Esp given every machine I have these days has >10 CPU threads and it won't be long before 100+ is normal.

Interesting! I wrote this approach up (more or less - extract design system -> make templates -> export) some time ago and I've found it unbelievably powerful: https://martinalderson.com/posts/how-to-make-great-looking-c....

I use it all day every day with Claude Code. I sometimes wonder past code if this has had the biggest impact on my day to day productivity, either having to make do with semi-bad looking reports or have a designer design them (which is slow).

Sort of feel sorry for Figma in a way though, given all the "partnerships" (highlighting their MCPs) and case studies they've done with Anthropic and then they release this. I note there isn't a testimonial from them this time.

I'm surprised how poorly Figma have used "AI" in general - given they were the "gold standard" in taking emerging technologies (WASM etc) and making an incredible product. The Figma Make thing was incredibly underwhelming, I managed to extract the system prompt out and it's basically just Gemini 3 Pro with a design prompt. Perhaps the original team has left?

They are extremely exposed imo. While all the UI/UX designers will continue using it for the forseeable, I strongly suspect a lot of their (A/M)RR was coming from extra seats for PMs, developers, etc to view and export and do commenting on the files - not core designer usage. I think a lot of this just won't happen on Figma as much.

I feel like I'm on a different planet when I see this kind of comment.

What if you need to call multiple external APIs at once with complex json? Sure you can call them one after another, but if each take (say) 2s to return (not uncommon IME), then you are in real trouble with only one thread - even if it is just for one "request".

I guess I'm spoilt in .NET with Task.WhenAll which makes it trivial to do this kind of stuff.

The regulation has an opt out for court orders though, which these are.

Yes of course but the spot market and associated futures has a very big impact on any "OTC" deal that a supplier and generator does.

Like you are not going to agree a eg 3 year supply deal with $SUPPLYCO at a significantly lower price than what you could get on the spot market for it (or what you could hedge out on futures).

UK peak power demand has _reduced_ by 15GW in the past 20ish years.

Obviously there is some investment needed but if you take a look at the capex cost of the north Scottish grid upgrades PLUS the HVDCs it's pretty terrifying.

The idea is that the financing will be "paid off" after 35 years so it doesn't require a "guaranteed" price after that (it reduces finance costs significantly when HMG is underwriting the main payback period. I expect the remaining 25 years will be extraordinarily profitable for EDF. Even if there is more maintenance costs they will have no finance costs. And finance is the main cost of HPC (60-70% goes to interest payments on the debt).

I'm certainly not saying we should stay on fossil fuels, my main point is that for the money that is being spent on various renewables could have been spent on a giant nuclear expansion at lower cost and far higher relability.

I don't really agree with this 'declining grid' narrative the renewable lobby has pushed. Yes there is upgrades to be done, etc etc. But peak UK electricity demand is down from ~65GW to ~45GW (which may change, but doesn't look to be).

Nearly all of the cost on the grid is to do with renewables, not 'general upgrades'. We would not be building 10GW of HVDC from scotland to england. We wouldn't be doing a drastic 275kV -> 400kV rerating and duplication in the middle of nowhere scotland otherwise.

Yes of course it's hit major delays, but my point is in hindsight had we started 5 nuclear plants (sort of as planned) around HPC construction start instead of spending the money on renewables, we'd have 5 plants coming online in (say) 2030. This would be nearly 50% of the entire UK supply. Another ~10 years later you'd have another 5, giving ~30GW of baseload.

And there's no reason we couldn't have used other reactor designs, apart from lack of financing so only EDF gambled everything on the EPR for HPC (again) and failed to deliver on time/on budget (again).

>Hinkley Point C has a contracted price of £133/MWh. I imagine there's some risk-sharing and inflation in the contract so this will probably go up. Compare this to £65/MWh for new solar and £72/MWh for new onshore wind [3]. Plus of course that wind and solar projects don't take 20 years to come online.

In 2012 prices HPC was £89/MWh. AR7 is delivering (in 2012 prices) offshore wind for ~£65/MWh.

But regardless these aren't comparable at all, for the reasons I set out above. That £65/MWh often doesn't include grid upgrades (which aren't required to nearly the same scale for nuclear, as they tend to be nearish existing population/demand centres and have existing grid infra). AND you still need (expensive) backup for that wind. We are building (right now!) new gas peaking plants that are allowed by law to only operate 10 days per year. The cost per MWh if you include that capex is horrendous.

>Lastly, the only way the per MWh costs can even get that low is by giving Hinkley Point C a 60 year lifespan to amortize the cost over a sufficiently large timespan. It's likely that as the plant ages, operational costs will significantly increase beyond what's projected.

Not true, HPCs 'agreement' is for 35 years. After that they just get the market price, so it is not based on the 60 year lifespan per se.

Australia and Canada have enormous uranium deposits. We're not going to have an entirely British energy supply chain (solar panels are mostly from China, etc).

And it's a relatively small amount of mass to ship - perhaps 500 tonnes of yellowcake for a reactors' yearly requirement.

I don't quite understand. The opposition can't be that insurmountable for offshore wind off England coast (instead of Scotland) given 700 giant offshore turbines are being built as we speak off the English coast. Including the world's largest windfarm (Hornsea 3).

The UK had/has the exact same windfall tax on energy:

https://www.gov.uk/government/publications/electricity-gener... https://www.gov.uk/government/publications/energy-profits-le...

And the reason Spain is so well insulated is because they have limited gas interconnection so they have a 'captive supplier' in Algerian gas. Algerian gas can basically only go to Spain, Morocco or the domestic Algerian market. They have some limited LNG export capacity (which is growing and will significantly change the price Spain pays longer term).

I mean offshore wind, not onshore. Thankfully most of the new ones are being built in England - check https://openinframap.org/#7.05/53.375/3.069, and these aren't curtailed.

But it was incredibly dumb to build many GW of offshore wind in Scotland when the grid was already over capacity.

It's worth thinking how this would work if it didn't work like this though? Nearly all 'commodity' markets clear this way.

If you switched to 'people get paid what they bid' it's almost certain the market would just converge back to this anyway - but with a lot more gaming and guesswork (wind guessing the gas marginal price to try and get the highest price).

The issue though is that the UK spends something along the lines of £10-25bn (depending how you count subsidies) a year on renewable. Something like £10bn in direct subsidies via CfD and RO, then another ~£1.5-2bn on curtailment (paying wind farms to turn off), ~£3bn in balancing market costs, and £6bn on transmission upgrade costs (passed back to consumers/businesses) to upgrade transmission for nearly always renewable projects. There's actually a lot of others I think you could attribute to renewables but I could spend all day writing about this.

My guess is that £20bn/year is a fair cost overall in subsidy payments. This is clearly not offset by natural gas fuel savings even with elevated prices.

The UK IMO made a couple of critical mistakes. Firstly, far too much offshore wind is in Scotland when it should have been closer to population centres in England. A few factors for this but the issue is planning is devolved to Scotland (so they have every incentive to approve as many) but energy subsidies are set by Westminster. By the time UK central government realised this it was too late (or they didn't want to rock the cart for political reasons post/during Scottish independence referendum).

We're now having to pay £20-30bn+ to get Scottish wind generation down to England where it is needed (primarily through new 5 (!) 2GW HVDCs from Scotland to England). It would have been far far better just to... build those wind farms closer to England. This would have still required grid upgrades but far cheaper ones (bringing it 100-200km to population centres instead of all the way from Scotland, plus you still need to do the ones in England on top of that for the most part to get it from the HVDC landing sites to the population centres).

The second major issue is there is definitely massive diminishing returns from adding more renewables at this point. There's too many renewables on the grid a lot of the time, even if transmission was perfect - supply is outstripping demand. Instead of building more and more generation the subsidies should be redirected towards storage projects.

But overall, for the same £20bn a year you could have probably built 5 Hinckley Point C sized 3.2GW nuclear plants concurrently (assuming £4bn a year capex for 10 years). In 20 years you'd have probably 30GW of nuclear built, which should cover nearly all electricity demand in the UK in that time, with very limited transmission costs (existing nuclear plants have good grid connections and you build them close to them). And importantly, you would basically eliminate _any_ dependence on gas from the UK grid. Clearly nuclear has risks in project delivery, but at least it's reliable once built.

Well a GPON frame is 0.125ms. So the best jitter is ~0.125ms, but even if the entire upstream segment was saturated you'd be looking at probably 1-2ms at absolute worse case. The OLT will not allow one user to hammer the upstream badly like you get with DOCSIS.

So it's really a non issue (XGS-PON is even better as more data per frame means heavy upstream 'clears' the frames quicker), IME consumer routers add more jitter themselves even with ethernet on the LAN side (and WiFi is a lot worse).

Probably just the equipment is far far cheaper now (customer ONTs and telco OLTs) than point to point is my guess?

You're confusing XGS-PON with NG-PON2. NG-PON2 isn't really used anywhere AFIAK, the "real" upgrade path is 50G-PON (50G/50G shared). NG-PON2's tunable lasers are expensive so didn't get traction.

Yes but if you compare urban areas (where 80% of people live in both continents) in US and Europe it's not massively different (Europe maybe 2-4 more dense depending on the country/city).

Obviously you're not going to lay fibre to the last 1% of population in the US (for the most part).

That's actually terrible. The UK (which has been a laggard in FTTH in Europe until very recently) is at 84% FTTH coverage already.

I'ts almost certainly shared. 99% of FTTH is (X)G(S)-PON which shares the fibre over a few properties. Usually something like 32 max.

The Swiss use point to point fibre (there are a few small pockets of this elsewhere). But in reality it is very hard to saturate. XGSPON has 10G/10G shared between the node. GPON has 2.4gbit down/1.2gbit up shared across the node.

In reality point to point is not really a benefit in 99.99% of scenarios, residential internet use cannot saturate 10G/10G for long, even with many 'heavy' internet users (most users can't really get more than >1gig internally over WiFi to start with).

And if it is a problem there is now 50G-PON which can run side by side, so you just add more bandwidth that way.

Yes you're right technically, but in reality you'd be swapping them the (vast?) majority in and out per inference request so would create an enormous bottleneck for the use case the author is using for.

Just FYI, MoE doesn't really save (V)RAM. You still need all weights loaded in memory, it just means you consult less per forward pass. So it improves tok/s but not vram usage.