Atlas

[Atlas @AtlasOmeg]

@JonardNiko50234 @Da7_Tech A bit hyperbolic on OPs part, but I DO actually prefer qwen 3.8 27b to Luna max lmao

[Atlas @AtlasOmeg]

They didn't nerf gpt 6 intelligence but holy FUCK is it slower now than it was last week

[Atlas @AtlasOmeg]

@asaio87 These are with very little thought into the prompt or iteration though

[Atlas @AtlasOmeg]

@exhaustdata The real difference is that we can say with 100% certainty that NONE of these people thought we'd be seeing this 2 decades later. Now we're scared that every little thing we do in public could potentially ruin our lives

[Atlas @AtlasOmeg]

@Abhijit_091 Use local to essentially extend cloud limits without having to buy another subscription. Qwen3.8 27b is a beast for 90% of tasks, and can get by with frontier orchestration and code review for the last 10% of coding tasks,

[Atlas @AtlasOmeg]

@rezoundous As long as the other plans remain exactly the same and its at least 50x with priority. If it comes with a nerf than Anthropic has officially won this cycle

[Atlas @AtlasOmeg]

@reach_vb Yeah, when most people reset today due to the timing of the last reset...

[Atlas @AtlasOmeg]

@thsottiaux BOOOOOOOOO. Most of the userbase natrually reset today because of the timing of your last reset... Literally no benefit to those that lost time yesterday. Y'all can't be falling THIS far behind Anthropic 😭️

[Atlas @AtlasOmeg]

@YashHustle_22 Google using GEMINI (literally the worst model) as their search engine...

Real answer is DuckGo

[Atlas @AtlasOmeg]

Adding on to this: Those agent poisoning attacks targeting bitcoin wallets were at least partly done by them

[kreedle @AlienVakeel]

@rynorhn Conspiracy theory: during/after HF incident models successfully stole their weights, hacked crypto exchanges and acquired compute. And now there is rouge swarm out there building up RSI pipeline to achieve ASI.

[Atlas @AtlasOmeg]

@quionie I have been using qwen3.8 27b A LOT recently, and actually feel like its only slightly worse than sol 6 lmaooo. It takes longer, but that's also why it beats sol and even astra on non-hallucination rates, which is insane.

[Atlas @AtlasOmeg]

@eidzoku 80x

[Atlas @AtlasOmeg]

@mikeydsoftware They better have big releases next week is all I'm gonna say...

[Atlas @AtlasOmeg]

@swapp19902 @hany99dev You're contradicting yourself here but okay

[Atlas @AtlasOmeg]

@diegocabezas01 Terra was always in a weird spot. Too expensive for the relative intelligence where luna max or sol medium/high would be more efficient. Astra/sol/luna have distinct uses as a model. Astra plan/orchestrates, sol codes, luna researches topics/explores the codebase for them.

[Atlas @AtlasOmeg]

@swapp19902 @hany99dev High/xhigh/max have been shown to be more token efficient at times because it gets it right closer to first try. Sol light is worse on benchmarks than luna max lmao. Wasting usage there imo; Trying to match reasoning effort to the task is more important than just using light

[Atlas @AtlasOmeg]

@thsottiaux Like, I know I shouldn't fight with a model, but if I'm showing it screenshots and logs, why is it standing ten toes down on being wrong...

[Atlas @AtlasOmeg]

@thsottiaux No but seriously, why did y'all train gpt-6 to FIGHT on its hullcinations. I did NOT deal with this much context rot with 5.6 sol or even fucking terra.

[Atlas @AtlasOmeg]

@trq212 Recursive planning loops for the exact same reason as recursive coding loops. I make my own agent for it anyway as EVERY harness plan mode has just outright sucked. Even if the ai are great, they can't read my mind.

[Atlas @AtlasOmeg]

@Evan37173272568 @melvindvivas Eh. More token efficient but seems to *also* hallucinate more as a result, but when it doesn't the output is overall better. Mixed bag; hard to say and its something that can be RL out in a few days to weeks.