OnPremConnor

Local AI Enthusiast | Dev | Student

[Justine Moore @venturetwins]

Sydney Sweeney reveals in an interview that she doesn’t care about the PHP vulnerability because she moved everything to Rust in 2024.

“They said the rewrite would take six months. We’re on month 30. We haven’t shipped a feature since, but I’m having a very normal Wednesday.”

[OnPremConnor @OnPremConnor]

@skyfallai When is the release ? Tired of seeing outputs of a model I cannot access. Would love to try it.

[Sam Pasupalak @spisallyouneed]

Introducing SKY-001, the first Engineering World Model.
On CAD and BIM tasks, it beats GPT-6 Astra and every frontier model we tested:

-> 51-125x faster
- > 150-210x cheaper
-> Higher IOU than all frontier models

SKY-001 lets you dream a physical object before you build it.

18 months of work. Built by the team out of Maluuba's deep learning lab (acq. by @MSRResearch).

The first ‘AI General Engineer’ will be the backbone for Autonomous Manufacturing.

[OnPremConnor @OnPremConnor]

@Blackfrost_AI @ornith_ Hasnt it been out less than a few hours ? What did you like so much about it? Wondering why I would bother switching from qwen.

[OnPremConnor @OnPremConnor]

I'm not even close to a reputable source, just a student, but I'd say we are at the point of just about anything being able to be done locally. The more frontier the task only means more time spent by you "context engineering" and "architecting your agentic workflow".

Any local model at this point (example Qwen-3.8-9B) can be trained on your specific "But I need Opus 5.5 for that" task if you collect the data around it, such as successful or unsuccesful runs by other agents on the same work same tool definitions blah blah. Also with Jev now, the context management/architecture scene is soooo new and game changed. Inference work lighter almost everywhere. The evolution is seemingly never stopping, its wild. Qwen4 weights allegedly drop this month.

Opus 5.5 and any frontier multi-trillion parameter models are so big (and all that other concepts I dont understand yet), that they can work with a shittier harness or tool definitions (Claude Code, Codex), and they'll still more times than not give you high quality output. The tradeoff for frontier is frontier subscription hitting that bank statement every month, although you didnt have to spend time architecting. While local is your time + effort + selling a kidney for hardware. After the Opus 4.8 to Opus 5 model update this summer destroying Claude's behavior with the CC harness, I had to sell the kidney.

And for forming opinions on local models, the "hit or miss" you're describing is the "holy shit everyone on twitter is using <insert new local model> to do xyz and sometimes its dog shit and sometimes its the best!" ......... is all just difference between authors harnesses/quants/vllm flags/tool definitions/system prompts/user prompts, etc. You have literally zero clue what their setup is. Of course the outputs are hit or miss across people trying it out, they arent confined by Claude within Claude Code. The missing details that I feel are soooo importnat to a post like this are not just 0xSero either, I really enjoy his posts. Its genuinely just everyone. Usually no user prompt even given, only shown output. Maybe I'm missing something tho, ofc wouldnt be first time.

Now is the time to go all in. Can't seem to find much to contradict that.

[OnPremConnor @OnPremConnor]

Wouldnt your Astra/Opus session just be the one to put a stop to it ? Thinking if I tell Astra to

"Look at this server on my local network and try to gain access to it without a password"

No way Astra properly writes out high value prompts to the local models and puts all the work together for you without guardrailing? Or am I missing something?

[OnPremConnor @OnPremConnor]

@sudoingX One spark

[OnPremConnor @OnPremConnor]

@2Halves1brain @sudoingX Holy shit nice ! Puts my 1 spark + 4080 to shame. I’m on my way when I can get sparks for 5k again !

[OnPremConnor @OnPremConnor]

@UselessDope @CollinRugg My exact thoughts. Makes me sick to my stomach, but all I can do from my place is hope the best for them.

[OnPremConnor @OnPremConnor]

@TheAhmadOsman Fire off a bunch of “autonomous” dots -&gt; codex tasks . That’s how I make them worth it.

[OnPremConnor @OnPremConnor]

@WescheNex1q Thanks for providing the prompt! Curious as to your top 3 models fitting a single Spark?

[OnPremConnor @OnPremConnor]

@bdvd_25 @StareN97 @Alibaba_Qwen Oh shit I have my 4080 system about a 3 hour drive away. Full thoughts on it so far since this ?

[lume.hl @LumeBTC]

Your only job should be to find the best way to be unemployed and still make money.

There is no other purpose in life.

[JK @GRAVEHOPP187]

You will spend $250 on a regular ass day in 2026 if you’re not careful

[kache @yacineMTB]

Don't use this
Build your own software using their tokens
So when their tokens become too expensive
You can switch to your own GPUs
Your compute. That you control
So you yourself can't be controlled

You need to control your stack -- or you become enslaved

[OnPremConnor @OnPremConnor]

@aijoey @Alibaba_Qwen @QwenDevs @NVIDIAAI 100% trying this but damn show me more than the aimless scroll! Thank you

[Dev Ed @developedbyed]

Ass Bench Update 🚨

After running GPT 6 Astra in a loop for the past week, with a couple of steers a day, we have officially reached high-fidelity cheeks.

I’m going to let this thing run in a self-improving cycle for as long as I can and see where the ass technology ends up in a month.

Lots of different butt styles will be added, so no one is safe.

Not even you, Granny.

You’ll also be able to support the project and help fund the benching of other models like Fable, Gemini, etc.

All it takes is a buck, and you can slap your sticker + website directly onto a cheek.

The future of AI benchmarking is here

More Soon

[Theo - t3.gg @theo]

@davepl1968 @DanielleFong Verifying the work, mostly

[OnPremConnor @OnPremConnor]

@Tech2Wild Great work ! And thanks so much for all the harness/prompt info. Followed.