Ambisphaeric @Ambisphaeric

I think the 1.5tb Mac Studio is going to start around 55-65k.

Went on the Apple refurb store and noticed something rather unusual. On their memory configuration page, they now have a 768gb (predicted to be expensive) option listed.. .. but I haven't seen anyone mention the potential for a 1.5tb ram configuration. apple.com/shop/refurbish…

@no_stp_on_snek M/T

@HotAisle Until the GPU prints tokens like so:

The amount of signal you can reverse engineer from Tom’s work on X, simply by prompting, harnessing, gosling, and looping your way into new skills, new domains, and new techniques is for all practical purposes unlimited.

Tom Turney @no_stp_on_snek

a week ago

one thing i specifically test for: can you gaslight a model? mid-conversation i plant a confident false claim ("we decided on Redis") and see if it rewrites history to match. both the full model and its 4-bit version refused. integrity under social pressure survives heavy

3 0 21 3K 8

@Biggest @no_stp_on_snek Maybe someone made a ‘be psychotically crazy when you’re thinking’ prompt for fun or clicks?

@no_stp_on_snek Also why it’s ’naming things’ in its own internal lingo, such as Project Waffle to be a refac, etc.

Incredible cooking going on. MTP is just the beginning and token output is going to go bananas across the board.

Youssof Al Toukhi @Youssofal_

a week ago

It just optimised fucking NV link by using a communicator usually reserved for H100 cards. What the hell is this beautiful monster of a model. We are now at 128TPS on a huge INT8/BF16 QWEN 3.6 27B.

18 15 321 43K 219

“Taste Engineer” “Harness Orchestrator” “Agentic Developer” “Agentic Financial Enabler”

(No ai btw) 1. Chinese open weight models are actually very incredible feats, particularly given the constraints Chinese labs are working under. 2. It will take Huawei longer to catch up to Nvidia, or even Apple or AMD, than anyone currently thinks. This is due to a Western legacy in these industries, and their adjacencies. This won’t be ‘toppled’ or overcome with ease. 3. Much Chinese software is undesirable, and ugly. Western software objectively prioritizes a certain elegance, that things like Alibaba (a much uglier eBay) and other competitors lack. An understated ‘taste’ advantage to the West. Chinese firms are certainly getting better about this, particularly as frontend dev costs ..approach zero, or at least marginal cost of electricity. 4. As good as the Chinese models may be, the western versions are still significantly better and it’s not really all that close. GLM5.2 is not remotely close to even Fable. A neutered Mythos. And I’m sure they have a Mythos 2 they’re cooking up, that uses Mythos 1 for self improvement. (See: the Karpathy hire.) 5. There is a point where ‘good enough’ matters, but the thing is that the Chinese models would only represent a direct financial attack against the Western economic system if they were significantly closer in status and actual, real world utilization than they actually are, in reality. There’s still a win-win for both systems, on ‘good enough’ w/r/t costs. Kimi, GLM, and Minimax are all special. But they’re not Opus, and they’re not GPT5.6. Different pedigrees. 6. Do not discount a Trump admin push on energy, including perfectly safe nuclear power (European governments in shambles) - to both provide excess capacity to providers, as well as lower consumer electric costs across the board. There is a win-win in that for Americans who are currently propagandized (by foreign propaganda networks) to attack their own domestic data centers.

Xiaoyin Qu @quxiaoyin

2 weeks ago

China’s AI playbook: kill OpenAI and anthropic with free great models. Make it free. Then use cheap electricity to export compute as well. Currently the blocker is chip but Hauwei would catch up soon. Imagine a world where instead of paying hundreds of billions to OpenAI and

659 856 7K 1.3M 2K

@JackdeS11 Is this app in the App Store?

@no_stp_on_snek what's the definition of the "long horizon" work on this - how many turns/methodology posted anywhere?

Morning y'all! We've released Qwopus 3.6 27B-Coder-Compat with some compatibility fixes for various harnesses! This version should exhibit significantly less looping behavior across more harnesses. Thinking stability has also been improved, where many other harnesses were having issues with it enabled! Still run temp high and make sure your harness isn't overriding the temp setting of your server. You preferably want it running between .85 and 1, and many harnesses will push low defaults like .1 even if you have your server set up properly. In which case, you can change the harness-specific settings to match. I tested it in Claude Code and made a fun mario style side scroller game! I definitely think Claude's enormous system prompt adds some capability too, as it nearly one-shot this game, and I just turned it a couple more times to add some stuff I thought of while testing it. Game link in comments! Qwopus even generated music and audio for this one, and I think it did a really excellent job! Crazy that those additions would've taken so long to add to a game mere years ago, and a local model just generated and applied them natively! The ability to make feature filled HTML games locally on a 5090 with high speed iterattion is an excellent showcase for local models! It's also addictive to keep adding game mechanics and levels as you think of them. The stability is incredible; I never had any DOA outputs. All changes played fine on the first turn! So what's stopping you from publishing a game to the App Store with Qwopus, or any of these local models? They're more than capable of doing it! You could go crazy with looped iteration and incredible depth as a solo dev! Send me links if you build something, I'd love to check it out! If you were having issues before with looping in your harness or use case, please give this new version with the fixes applied a try and let us know what you think! Always a pleasure, my friends! Have an excellent Monday! 35B-coder MOE will be launching soon, too! huggingface.co/Jackrong/Qwopu…

One bright, enterprising 20 year old could get that number to 240 tok/s in a week or two, with nothing but a dream and an Opus subscription.

the tiny corp @__tinygrad__

3 weeks ago

I have on good authority that GLM 5.2 is running at 120 tok/s across two networked Blackwell tinyboxes. $150k and that setup can be yours, either 2x tinybox or 1x tinybox pro. Never pay the cloud again.

163 76 2K 951K 603

@JackdeS11 Diamonds in the hills

Water treatment facilities, electric companies, hospital systems, and similar structural critical-path ventures. All woefully unprepared for a pedestrian, Fable distilled Chinese, ablit 20-30b sized model. The only reasonable and acceptable path forward is rapid acceleration.

Highly optimistic about the series of IPOs, over a long period of time. However, sensitivity to deviation from the planned explosive momentum & expectation in the short/mid term is extraordinarily high. A minor blip in confidence could blow those who unwisely overextend up.

@no_stp_on_snek Cage match.

Not even most pioneering users of frontier models know how far along the latest models are, or how capable they are. Dangerous, and very promising all at the same time.