Civic Intel

God bless America

[Civic Intel @CivicIntelMain]

It's still not great... and any model can use a skill - the quality of the skill use will scale with the quality of the model, right?

[Adam Holter @AdamHoltererer]

GPT-6.1 Terra xhigh Frontend Test:
Without Skills vs. With Skills

[Civic Intel @CivicIntelMain]

"with more usage" yeah ok bro. Telling on yourself

[i²cjak @i2cjak]

no that's sick man yeah 27B locally cool yeah you spent $10k? so I pay $200/mo and get access to a model 10x better than that with more usage so that's really cool though yeah

[Civic Intel @CivicIntelMain]

@0xSero @incaseyouf0rgot People who call things cringe haven't built anything themselves. They just critique others. Ignore them.

[成龍 @graphtraversal]

at walmart rn and everyone looks abliterated and quantized

[Civic Intel @CivicIntelMain]

@MiaAI_lab Wow, that bad? The htmls you posted were pretty rudimentary. The model gave their design no flair or thought at all.

[Civic Intel @CivicIntelMain]

@gabu3d_pl I did the same thing and I’m back on it. It’s too buggy to use out of the box but after having codex tweak it, it’s finally reliable enough. I’m on qwen 3.8 flash next

[Civic Intel @CivicIntelMain]

Why is it always Minnesota, man

[Daily Loud @DailyLoud]

Registered sex offender, 29, busted trying to enroll in high school as teen girl again 👀

[Civic Intel @CivicIntelMain]

I got fed up with the model at capacity warnings and switched back to 5.6 Sol. It’s just how I remembered it, and cuts out a lot of the points of frustration that I had with 6 sol and 6.1 sol.

[Benjamin Marie @bnjmn_marie]

I really tried GPT 6.1 Sol (with Ultra in Codex), but it's just not close enough to GPT 6 Astra.

I don't know how OpenAI could conclude that they perform similarly. That's just not true.

[Civic Intel @CivicIntelMain]

@fraserpricee @itsjustmarky Legend

[Civic Intel @CivicIntelMain]

@seanhighness No, the gguf.

[Civic Intel @CivicIntelMain]

Chinese twitter has convinced me they’re insane

[Civic Intel @CivicIntelMain]

Indians are AGI

[Austen Allred @Austen]

Oh my goodness this has to be right.

Instead of generating a random 256-bit number they generated a random number between 0 and 256.

That’s how they got 4, 66, 155, 156, 230 and 237.

[Civic Intel @CivicIntelMain]

@loktar00 I heard next week!

[Civic Intel @CivicIntelMain]

@0xSero Just awesome. It will always impress me to see a high quality model run at high speeds. It’s just mind blowing

[Civic Intel @CivicIntelMain]

@catonooka Beautiful. Those are great results!

[Civic Intel @CivicIntelMain]

@Aries @FinanceDotDev Let us have paper trading accounts! Alpaca is currently the best at this but they only allow a max of 3. Super inconvenient

[R A W S A L E R T S @rawsalerts]

🚨 #BREAKING: Virginia man charged with duct taping his wife to a chair so he could watch football upstairs “in peace”

[Civic Intel @CivicIntelMain]

It highlights such a key problem with reported output speeds- the model quality is neutered to get those, and the people reporting the speeds just hide or don’t address it. I wish we had a comprehensive judge of quality that accounts for kld, top-1, output speed, prefill speed, and size on vram/ram. Just take everything people need to know and put them together so people don’t need to guess. Maybe I’ll make a site like that.

[まぷて @mapte_fireball]

@FreeofspeechUSA The numbers come from the Basalt developer's benchmarks, not my own testing. They report up to 2.6x vs Strata, though results vary by hardware and quantization.

https://t.co/JaZS7KdKJb