Frontier models for frontier problems, but what can you do if you let a surface laptop ultra dream all night?

Disappointing it only did a security review and created some prs... I thought you were going to say it had cured cancer or something 🤷🏻‍♂️

It’s a laptop

It's a joke

LOL

Impressive technical demonstration, but at around £6,000 in the UK, the price is hard to justify. A £1,000 laptop with a Codex or Claude subscription offers better value, access to more capable models and continuous improvements. The only real advantage I see in running AI locally is privacy. Beyond that, I’m struggling to see the justification for the hardware investment.

Very cool. I feel a little bad for the poor guy working all nighters. *Slap* Stop anthropomorphizing, meat proxy.

Interested in your experiments in fine-tuning

for sure

For some scenarios, there is a clear advantage of having a model that doesn’t depend on the internet connection. Natural disasters? Rescue missions? Regarding dreams: "We are like the dreamer who dreams, and then lives inside the dream. But who is the dreamer?"

Qwen 3.8 Flash Next is a much better fit for this machine than the dense Qwen 3.8 27B is FWIW. Makes better use of the shape of your machine: high VRAM, decent compute, but low bandwidth @ 300GB/s (versus e.g. 1700GB/s for a 5090).

thanks for the tip, will test now!

People are also largely using inference stacks that are model-and/or-device-specific these days, and seeing a lot of gains from that. E.g., this might be one of the inference stacks you might run that model with on a Macbook:

Do Ultra laptops Dream of Electric Sheep? 😉

Beyonce haha 🫡💯

Brilliant content. Thank you. Your office is looking like an Aladdins cave of tech. I for one would love a product/tech list!

@grok calculate how much 21.3M Input tokens (19.6M cached) / 233.3k output tokens cost on DeepSeek 4.1 Flash

Hey Scott, what model Mac mini would you recommend for overnight agent work? I’ve heard they are good for this kind of thing

This is awesome. Is it genuinely the same power usage as a lightbulb, at least cost wise? What ultra is this? And how long would you say it takes for the laptop to pay for itself if overnight background tasks were the only thing you’re using it for?

In this case, I’m talking about 100 watt incandescent bulb, not an LED bulb, so that’s kind of a bit of a joke, but it’s 100 watts, it’s not like running a hairdryer or something. On a pure tokenomics perspective, it probably would take a while for it to pay for itself because cloud models keep getting cheaper and cheaper. A more interesting way to measure it paying for itself would be if you were renting a GPU in the cloud for eight or $10 an hour versus using one locally to do something like fine tuning, it could pay for itself in less than 1000 hours of hard use. That would be a more realistic measurement

Which model are u running locally Scott

While I don’t like to anthropomorphize AI, you can see the exact moment the agent escapes the GPU and takes control of Scott.

can you share like how you setup i like to try on my laptop may be a small model but want to try it and leave it to run want to see what it find out Also how you train the model for particular project or task or what and how you train it

What’s the software you are using?

GitHub copilot app pointed to Nvidia pair hosting Llamacpp

Thank you 🙏

My Surface Ultra is on the way, and I'm going to try to fine tune a decision model