ac29

I've had seasonal allergies for decades and haven't seen them trend more intense, though some years have more allergens than others.

Personally I only take allergy medication maybe 50-100 days a year, and usually just a half dose. I have definitely heard from people with worse symptoms that they get a tolerance to medications so it may help to switch between them if you take them year round.

> In the US all transmitting at 1W with these radios is contrary to regulations.

With a ham license the limit is 1.5kW, without one its zero

VHF doesnt need line of sight, it has excellent ability to penetrate obstructions due to its long wavelength. For example, its not difficult to receive FM radio transmissions indoors, even from tens of km away. Some obstacles will effectively block any radio signal though, such as solid earth or concrete (esp rebar reinforced concrete).

You are right that handheld radios wont get more than about 10km, but that is due to the curvature of the earth. Mountain top to mountain top, you could easily do 50-100km

Just to be clear, "Plus" and "Max" Qwen models are closed. Seems likely smaller open versions will be released, but that's not what was announced today

> When an agent hits an unfamiliar codebase it typically does "cat file" or reads the whole thing first, at least in my experience.

Depends on the size of the project and specific files. I have definitely seen agents make smart use of pi's "read" tool, which can take an offset and line limit (or defaults to a max 2000 lines/50KiB if the model doesn't specify). The bash tool also has the same max output, so if a model decides to cat instead of using the read tool it still wont blow out its context window with a single large file read.

But this sort of thing is going to vary with harness, model, project, and whatever the RNG delivers for the day.

Doesnt "a bunch of SRAM" top out at maybe a few gigs per chip (with zero area used for logic)? You'd need an order of magnitude more to fit even a fairly weak general purpose LLM model.

I've never seen a charger in the US that doesnt operate 24/7 regardless of grid load, is this different in other countries?

> Microsoft has ensured the alternative is nearly impossible, constantly working to block any workarounds that users discover to use a local-only account.

Local accounts still work fine for Win 11 Pro, I installed it a few days ago using a clean ISO directly from Microsoft. No special patching or command line stuff needed, making a local account is part of the official install process.

> With CO2 emitting option is priced out or banned

GP was talking about injecting the CO2 back into the well, not releasing it to the environment. There are even standards for specific injection wells used for long term storage (EPA Class VI).

FYI, distilling Gemini is explicitly against the ToS:

"You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio). You also may not attempt to reverse engineer, extract or replicate any component of the Services, including the underlying data or models (e.g., parameter weights)."

> ChromeOS meanwhile has the worst compatibility off all four

ChromeOS can run desktop Linux software and Android software, so it definitely isnt worse than Mac. Its probably even better than Windows. Of course, if you need Mac/Windows software, Web/Android/Linux alternatives might not exist or might be worse. But the devices are hardly lacking software compatibility.

Its horrible with memory, launching a single empty terminal uses 307MiB on my Linux system

> Enormous numbers of consumers own $50,000 cars, but a pair of $2000 GPUs is "not consumer"?

$50k is a median priced car in the US. I'd guess >99.9% of people do not own $4000 of GPUs. I consider myself a computer person and I dont think I even own $4000 of computer hardware in total

> The range of these lora nodes is a bit of a myth. It is better than higher frequencies but you shouldn't expect anything more than a km with obstructions, realistically half that.

I've done a number of projects with commercial radios operating in the 902-928MHz unlicensed band and typically we target 1-10 mi (roughly 1-20km). Elevated antennas with enough gain to get you to the legal limit (4W EIRP) can get you a heck of a lot of range, even without line of sight.

With line of sight, communication to the horizon is possible.

If you're talking about the EU 869 MHz unlicensed band used by LoraWAN, thats quite a bit different and I'm less familiar.

> Has canned fruit actually lost popularity?

Compared to Del Monte's heyday in the previous century? Absolutely.

A remarkable amount of fruit is available all year, or most of the year now. I cant imagine eating canned fruit by choice.

Unsloth Dynamic, just some branding from Unsloth for their quants (other people use similar techniques)

Memory and compute/energy overhead

Not a great example since using Anthropic subscriptions with third party applications was never allowed, they just didnt take steps to prevent it until recently.

> My set up is not the best but it is working for me and is saving me money.

I've got a local setup too but unless you consider hardware zero cost, there is really no way to save money. The class of model you can run on <$5k of hardware is dirt cheap to run in the cloud (generating tokens 24/7 non-stop is a few dollars a day at most, possibly even less than the cost of electricity to do it at home).

Modern GPUs aren't optimized for MoEs though?

The advantage to a dense model like this Mistral one is that it is as smart as a much larger MoE model so it can fit on less GPUs. The tradeoff is that it is much slower since it has to read 100% of its weights for every token, MoE models typically only read about a tenth (though sparsity levels vary).

Payment processing likely eats up at least 2-3% of that

Thunderbolt 4 and 5 are just USB (40, 80 Gbps) with mandatory support for otherwise optional USB-C features like video and high power.

According to wikipedia the current marketing names for USB are just their speed: USB 5/10/20/40/80 Gbps. No version numbers or anything else.

Your table doesn't indicate reasoning vs non-reasoning, or reasoning level

Your benchmark has Opus 4.7 performing significantly worse than Sonnet 4.6. Even if true on your benchmark, that is not representative of the overall performance of the models.

There are Yoga rooms in terminals 1, 2, and 3

I'm using it via OpenCode Go, which claims to only use Zero Data Retention providers.

How much you trust any particular provider's claim to not retain data is subjective though.

The memory requirements aren't that intense. You can run useful (not frontier) models on a $2-5K machine at reasonable speeds. The capabilities of Qwen3.6 27B or 35B-A3B are dramatically better than what was available even a few months ago.

Practical? Maybe not (unless you highly value privacy) because you can get better models and better performance with cheap API access or even cheaper subscriptions. As you said, this may indefinitely be the case.

It was a weird point to make in the post given that exe.dev charges $0.07/GB for transfer. That's arguably worse than the major clouds, who charge about the same for egress but give you free ingress.

Google's naming might be misleading, currently 3.1 flash image outperforms the available pro version (3.0 pro) on most benchmarks: https://deepmind.google/models/model-cards/gemini-3-1-flash-...