The picture shows two power supplies. Powering what is effectively one appliance from different circuits is a definite no-no, and I can't think of any circumstance where it wouldn't be in a home.

If his mains supply was sufficient to run the server and the house in the first place then the simplest solution would be to simply upgrade one of the MCBs/RCBOs on one of the circuits to the required capacity. I am not sure a landlord would even notice something like that, and if the house is wired correctly in the first place, it's unlikely to be dangerous. So going from say, 6A to 12A, on a 20A mains supply is generally fine if the gauge of wiring is correct.

By contrast, my observation is that MOST of the world's population still works from about an hour after dawn until early afternoon, or sometimes until dusk, depending on their age, job, station in life and the general level of resources they have versus what they need. And they probably always did.

Basically this. They way we usually put it is that we want clients to "get on and off the channel as quickly as possible". That requires all clients in range of each other to be behaving (respecting the rules) and using fast enough data rates to minimise their consumption of precious air-time.

Under the hood though, it's a very granular frame-by-frame, almost nanosecond-by-nanosecond thing that leads to the overall throughput at a human timescale. To give you a sense, let me try to summarise the factors affecting throughput this way:

- Data Rate: the transmitting client can adjust the data rate of each frame up or down per frame if they want. For example, a single TCP session on a 2.4GHz channel could in theory see data rates everywhere between 1Mbps and 450Mbps. But in practice most drivers I've seen adjust up or down incrementally. And in a healthy network, they usually hover around the top 25% of the mutually supported data rates (but they also spend very little time at the highest data rate, typically less than 1%). Also the AP could be using different data rate to the client, and usually is. The rx and tx directions are effectively separate streams and data rate is always chosen solely by the transmitter.

- Block Size: Similar to TCP windowing. Data can be sent in multi-frame 'bursts' before an acknowledgement is required by the transmitter for it send more. In the original Wi-Fi, every frame had to be acknowledged. Later standards introduced this idea of block acknowledgements.

- Re-transmits: Whenever acknowledgements are not received, the data has to be resent. Block size will be reduced, possibly to 1, so it will also take longer. Note that re-transmits are expected and very routine in Wi-Fi, whereas in TCP they are usually considered more of an exception (except on the internet). I've observed re-transmit rates of 20% in networks where no user is perceiving any sort of issue at all. So Wi-Fi is very robust to frame loss, up to a point, but even so, re-transmits do end up having a large impact on the aggregate throughput.

- Clear channel wait time: It's no exaggeration to say that transmitters spend most of their _waiting_ to transmit. And a big chunk of that wait time is just waiting for the medium to be clear - the clear channel assessment. If the client thinks there is a transmission going on, it just has to kill time.

- Other wait times: Even when the channel seems clear, there are various requirements to do nothing before and after transmitting. For example, the inter-frame spacing interval and the random back-off interval. These are just the rules of play. In fact, congestion avoidance on Wi-Fi could be said to be entirely a matter of timing.

Note that these are a simplification and clearly I can't mention everything or cover all the nuances. But, in the way I've framed it here, the clear-channel wait time and the re-transmit rate do basically encapsulate the impact of intangibles I didn't mention, like congestion and noise/interference.

TLDR; Wi-Fi transmissions are extremely lumpy at their native timescale, but many seem a lot smoother than many TCP transmissions at human timescales.

> Correlated, but obviously bad code can really fuck with neighbors.

Also true. Bad code is usually exemplified in Wi-Fi by bad drivers (looking at you Broadcom). These will cause clients to "stick" to bad APs when they should roam, or pick the wrong channel/AP/band in the first place. Intel is generally very good.

> And each client has an incentive to be greedy so users of that client get a better experience.

Greed is good in the sense that clients want to transmit their data as soon and as fast as possible and we want them too! But they have to respect the rules. Of course there's only a handful of chipset vendors so they mostly do. But within that, there's still plenty of room for clients and APs to do things that are _sub-optimal_ even if they are Wi-Fi legal, as per the sticky client example I mentioned.

> So you fall back again to QOS for what you care about..

Wi-Fi does indeed have its own implementation of QoS which is of course a timing dance! But I think you're referring to QoS in higher layers like IP. So it's worth mentioning that this WiFi stuff is all happening at layers 1 & 2. All the congestion detection and re-transmissions and so on that may be happening in higher-layer protocols like TCP are happening _in addition_ to what is going on at the WiFi layers.

If you have a good connection and are successfully able to transmit packets to your AP at 600Mbps, and your neighbour has a poor connection and is transmitting at 6Mbps to his AP at that moment, you literally have to wait ~100 times as long for a free medium before you can attempt to transmit. And that's for every single frame. Then you have to hope his client is well-behaved enough not to transmit while you are transmitting. Otherwise you end up having to wait again and retransmit anyway.

You might not notice this with only 2 clients. It might be the difference between a 80MBps and a 50MBps download for example. But it decays exponentially with the number of clients.

When you and your neighbour _appear_ to be transmitting at the same time, each adapter is actually spending most of it's time waiting for a clear medium and for various backoff timers to expire before attempting to transmit.

"Appear as noise" is not defined for Wi-Fi adapters. There is only "I received a frame addressed to me and acknowledged it" or "I sent a frame and either did or didn't get an acknowledgement back from the receiver". Receivers do not know why they didn't receive a frame, or, if they received a corrupted frame, why it was corrupted. They just wait for a retransmit. Senders ordinarily wait a certain time to receive an acknowledgement, and if they don't, the start the transmit wait cycle again. But they often then reduce the data rate to increase the odds of a successful transmission.

I'm glossing over some complexity here, because there's a sender and receiver to consider, and each has a different view of the RF environment, but the point is always correct when all transmitters and receivers (lets say the 2 APs and each has 1 client) are in audible range of each other. And this is most of the time. Note that "audible range" (where the signal is such that the medium is deemed as busy by the adapter) is much larger than the "usable range" (where data can be transmitted at reasonable speeds). So transmitters create interference in a much larger area than they actually operate in.

That means your neighbour transmitting at 6Mbps to his AP will indeed degrade the performance of your client who wants to transmit at 600Mbps because your client has to wait ~100 times longer for a clear medium.

Each data rate in the standard uses a different encoding technique. "Faster" encoding techniques cram more data into a given transmission interval but require a higher signal to noise ratio to be received without error. Since SNR declines with distance you can have a rough idea at what distance from a transmitter you will be able to receive at what data rate.

However, people and vendors focus far too much on maximum throughput. I've seen data showing that even in the best conditions, clients spend about 1% of their time transmitting or receiving at the highest data rates. Because they are dynamically adjusting the data rate based on the perceived SNR.

Individual clients' peak throughput also works against _aggregate_ throughput when talking about wireless networks with multiple users. If you have 100 clients, do you want one to be able to dominate the others or everyone get a more or less equal share? These peak speeds assume configurations that I would never deploy in practice, because they favour individual users and cripple aggregate throughput - things like 160 MHz wide channels.

But the sticker speed is what sells..

But at a fundamental level, the channel space (~60 across all bands best case) is extremely limited but the potential growth in transmitters is unbounded. It's like a linear hack to an exponential problem. It seems to work at first, but under very high load conditions performance still degrades ever faster until it falls off a cliff. Then there's all sorts of complex dynamic behaviour like the hidden node problem to add to this, but it all boils down to needing air-time and SNR.

802.11 is in general a vast swag of cool tricks, and when enough ideas are thrown at a wall, many do end up sticking, but for the most part the benefits are cumulative. MIMO being one major exception.

only 1 transmitter at a time per channel - across all WLANs, yours and your neighbours, with no deterministic way to avoid collisions.

It's a shared medium and it's not even half duplex, unlike the dedicated full duplex you would typically get with an ethernet cable to a switch port.

The fact that Wi-Fi achieves what it does with this limitation, and how it co-ordinates the dance of multiple unknown clients using the same medium - and in the presence of other RF technologies to boot - is indeed an incredible technology story, but this achilles heel is the single most defining thing about Wi-Fi performance.

AI companies are trying to ride a growth wave where the income curve lags the expense curve by 1-2 years, and at the same time investing 10x their historical income on next year's projected demand.

Everyone is selling their API calls at a loss, because to capture the investment required to scale the business up and the costs down, you need to grow your market now (in relative and absolute terms). And history shows, that in big tech you often have winner-takes-all situations, or, at least a couple of big firms will dominate, and the others will die. That's where market share becomes a key strategic goal.

But to secure that, they also need to be building next year's compute now. And if their anticipated compute needs are 10x this year, they've got a serious funding problem, one that can only be filled by capital with an appropriate risk appetite. You can only get this high-risk capital when the potential payoff is even more enormous, or, when it's a smaller bite of a much bigger pie. Hence, MS putting into OpenAI and so on. But the investment needs are getting so big we are starting to see some pullback from more conservative sources, but also record deals from others.

Now say an AI company does get the capital they need to grow. Well, they've still got a very serious supply problem. RAM, GPUs, water, electricity etc. Hence why there's a lot of deals and cross-investment going on - everyone is trying to secure resources and lower their overall risk exposure while keeping a foot in every possible door, so they can switch alliances whenever it's expedient, and because collaboration also helps the overall market to grow.

This all explains to me why the industry _needs_ the hype. These companies can't exist without it, because the money they need to sink in, in order to even be around in 18 months, far outstrips all reasonable financial practices. So it's capitalism on steroids or nothing. If you believe the AI story, then to that extent, it's rational.

But note that nowhere in this scenario does it suggest the actual consumers will be getting a consistent product at a consistent price!!!

But, overall, the current AI pricing is completely unsustainable, across all AI companies, except via the exponential growth they are relying on. Dylan Patel did the most insightful analysis of this I've come across.. https://youtu.be/mDG_Hx3BSUE?si=nyJu4adwYCH1igbJ

I have been wanting to try CC with different models since Opus went downhill last month..

What limitations or issues have you noticed when using DeepSeek with Claude Code if any?