syntaxing

That was a fun read, I wonder if vscode would still load the script if you click “don’t trust author”? That being said, can you embed something yourself and push the commit?

What spec is your macbook? I want to run Deepseek V4 Flash but its too slow for agents on my Strix Halo.

I don’t think they meant exclusively Chinese models. Many companies including the one I work for uses big US model for most of the work but data sensitive ones are on prem open source ones.

I want something similar to this but for Ubiquiti. I don’t need anything fancy, just something that audits my home config and tell me if I’m doing something stupid, dangerous, or both.

I don’t know if the llama cpp implementation is wonky (and only supports the binary version) but it’s a lot slower than 35B-A3B @ Q4_KM + MTP with CPU offloading.

For those curious about their demo, I’m pretty sure it’s using Locally AI (iOS only) that lmstudio acquired/aquihired a couple months ago.

In certain Asian countries, medical education is 100% covered but you must work at a public hospital for X years. I honestly think it’s very fair. If tax payers full funds your education, it should be mandatory to work in public services for X amount of years.

The reasoning built into the models matter so much too. I recently swapped my Qwen3.6 27B to ThinkingLabs’ fine tune and it does what it publishes. I cut my token usage in half, which is a big deal since I only get ~20 TPS for token generation.

I know quite a few people who got shingles in their early 20s. One of their doctors didn’t believe she had shingles until the blisters formed. The vaccine can definitely help those younger than 50, dementia benefits or not. Some of them have permanent nerve damage after getting shingles.

I saw that but end of the day, the chips themselves don’t have hardware support for FP4. There’s smart ways around this limitation but it will never natively be close to true FP4 performance like MXFP4 and NVFP4 (happy to be proven wrong though).

How has it been and how’s the speed? I read online that the it’s ~200 TPS for PP and ~15 TPS for TG. Unfortunately for those speeds, it’s very very hard to use for agentic stuff.

I own one, I don’t feel like the RAM is a huge issue (of course I want 192GB to run something like DS4 Flash). The lack of FP4 and slow memory bandwidth is rough. NVFP4 support is such a huge advantage that I would recommend others to buy a DGX spark over a strix halo if you’re using it purely for AI. Strix halo works better for general computing.

Highly recommend lemonade server if you have a strix halo desktop. Been using Qwen3.6-35B @ Q_8 as my main driver and it’s been great with 60 TPS for generation. I occasionally use the 27B @ q6 but only get 20-25 TPS for generation with MTP.

Posted in another comment but you cannot just think about John Deere's balance sheet. A whole other industry would have existed without John Deere's intervention but they were able to capture a lot of the GROSS revenue due to it. You can look at other non high litigation capture industries like automotive.

Admittedly I have never worked in the agriculture industry, but I have been a mechanical engineer for multiple industries before I became a software engineer (a good 5 years I was in a position where I quoted customers). You really cannot imagine that out of the 44B gross revenue and 5B net, that a "non tiny fraction" was not related to right the repair? Collections of receivables + Proceeds from sales of equipment on operating leases is north of half of the 44B gross. How much of that gross would have not existed should there been a third party market to repair and service exist products? I honestly can't give a number but I doubt its "tiny". Look at the car industry, about 20% of the global revenue is aftermarket. You simply cannot naively think that "right to repair" only effects the service contracts. Theres aftermarket parts and 3rd party repair shops that COULD have been a bigger market without John Deere's anticompetive practices.

Not OP but I went through some data and John Deere makes 5B NET profit for the worse years. 10B for their best (only looking back 10 years). I wouldn’t be surprised these anticompetitive (as in anti “consumer”) has netted them north of 10B.

I have another strix halo that I got for half the price (before this price increase world wide). AMD making lemonade is one of the best reasons to get a strix halo. Lemonade + qwen3.6 35B MTP @ Q8_0 + anythingLLM (in docker) replaced 90%+ of my AI usage. And it’s fully local! Setting everything up took less than 3 hours total, including installing the OS

https://lemonade-server.ai/

This is a great idea and hope it works out, especially on shoring chips back here in the states. That being said, their website is absolutely atrocious. One of the very few sites I got motion sickness from scrolling.

[1] https://fab2.com/

Fluid mechanics was my jam before my software engineer days! Unfortunately, I think the FAA strict regulations will prevent any of the good innovations. But you can do this thing called flow control where you can control where the boundary layer detaches on the wing. You can achieve different L/D ratios and save a TON of fuel cost. NASA supposedly was going to try it on the X-37B but civilians never knew if they did or not.

Not affiliated but I had good experience with GL.inet’s comet line [1]. They have one on kickstarter that’s the size of a google cast puck that uses purely usb-C. Though all my KVM do not have internet access (blocked at my gateway). I can only access it via tailscale externally.

[1] https://www.gl-inet.com/products/gl-rm1/

This is simply not true and a misrepresentation of the issue. There is no issue with SP500 buying into all these companies. The issue is that “valuation” should be determined by the market before the index funds buy into it hence the original rules and policy in place. Else wise we run into the issue now where our 401K is pumping the IPO, regardless of fundamentals.

Moral issues aside, this is a pretty clever way to get a wide variety of training data

What about self host Chinese models?

Hot take, barring from special edge cases, I find using dumber models (like local Qwen 3.6) to be the best balance. Smart enough to do stuff but dumb enough where I don’t trust it and verify what it’s doing rather than letting it do the third whole code base refactoring of the day. Also forces me to know my code base and ask very descriptive tasks rather than go “something is wrong, fix it”.

But worse data retention and training policy. It’s a balance.

If you look at openrouter for Deepseek model providers that does not use your data to train, it’s still significantly cheaper than Anthropic pricing. The Pro and Flash performs closely to Opus 4.5 and Sonnet 4.6 respectively (though no vision capability which is a fair thing another user called out). The pricing of Pro is close to Haiku. The pricing of Flash is 10X cheaper. To put into perspective, you can have Sonnet 4.6 capabilities at 10X cheaper than Haiku even without “Chinese government subsidies”.

It’s wild. Regardless of Deepseek direct pricing, on Openrouter itself, the pricing for Pro is comparable to Haiku. Flash is even cheaper. You get Opus 4.5 and better than Sonnet 4.6 performance.

Serious question, I have a Bambulab printer now. What is a good next printer option? Prusa is way too expensive without a decent AMS alternative. Flashforge is doing some sketchy stuff with their maker site + AI. Is Creality or QIDI the best next options?

Any brand of test kit you recommend?

Is there something that is good to be a “android” server? I want to sign in to this server for all my chat stuff and use beeper to connect to it. I tried using a tablet but the battery keeps dying.