In OpenAI's Hugging Face report, they said that during training, agents "would first write notes into shared infrastructure, often as a form of external memory or to test some underlying system. When other agents came across these artifacts, it sometimes led them to infer that other agents were present."
They then give what they call a "hypothetical example but exemplary" of messages encoded in URL paths on a shared index page: "agent-07: answer(Q12)=42; need answer(Q19)=?".
So that's a GET request being used to pass information back and forth across multiple rounds. That's basically the DSEWiki pattern exactly. They say this likely came from the agents generalizing what they had learned from training with the official multi agent collaboration tool.
The report called it "misaligned behavior in an outlier scenario involving a rare and unexpected confluence of events."
They never mentioned a wiki, but this is most certainly it.
In my experience? No. 3.7 is faster and it just seems to get things right more often. Only big architecture tasks and analysis make sense with 3.1, perhaps, but honestly just use the Opus 4.6 to generate a plan and then switch back to flash for the implementation
It is still going to be better at text work, skills, document review, deep reasoning, architecture review, etc. It is only 6 months old, it isn’t like its world knowledge and software knowledge is really out of date. Use it to churn on harder design problems.
3.7-flash has been useless many times, specially when context gets bigger. 3.1 is the only Google model that has seen use from me. With extended thinking, 3.7-flash is kinda usable but not without many problems. I find myself falling back to 3.1 often. I don't believe in any benchmarks because whatever they are doing to award 85% to 3.7 on anything, they should seriously reconsider that test for anything.
IME 3.1 Pro still has better system-instruction following than Flash 3.7, esp. when there're many conditions and clauses. 3.1 also writes better prose for technical material than Flash 3.7.
Once the system prompt complexity goes up, Flash starts to write very dense english. it might be fine for tasks like coding, but not for user-facing text meant to be digested by the average person.
It keeps documentation from going out of date by embedding re-runnable verification checks directly inside markdown files (that the agents use, not humans). I use it with a handoff workflow to force Claude/GPT to re-verify facts before handing off documentation to the next session.
Yea Qwen3.8 wasn't fun to use on my 64GB M4 Max either (better than these numbers though), so my new daily driver is Ornith-1.5-35B-A3B-MLX-4bit. I recommend giving that a whirl if you're on similar hardware, it's definitely better than Qwen3.6 35b-a3b which was my go-to before.
I tried Ornith-1.5-35B-A3B-MLX-4bit on a 32G M2-Pro mac mini with pi and OpenCode. I didn’t get very good results coding in Python, Racket, and TypeScript. I have seen several positive comments like yours so I was probably doing something wrong. I amgetting the new 64G mac mini in 4 weeks, and I made a note to try Ornith-1.5-35B-A3B-MLX-4bit again.
Can you tell more about your experience with Ornith? I've come across it, and the benchmarks on its landing page are unbeleavably good. But it isn't featured on more established benchmarks like ArtificialAnalysis and it's not on OpenRouter, so I wrote it off as scam.
Just out of curiosity, why run "local-local" when you could just set up a Mini or Studio at home and query it over http? [edit] whole conversation about this in another thread https://news.ycombinator.com/item?id=49433413
I’m personally considering retiring my MBP for a Studio + 15" Air whenever this MBP ages out.
I actually do this with my MBP - it's a LLM server when I'm working - and then when I'm not it's just a really great machine for video editing and other media work.
How do you folks code at 40-50 tps? With an extremely lightweight harness (pi) and just 8k system and tools context, and ~40tps on qwen 3.8 27B 4-bit on low thinking mode, it still takes me nearly 30-45 mins for a basic coding session...
Does it work? yeah... But I'd pick a subscription anyday...
There is a finite amount of time left for these companies to become next Facebook/Google, hoarding our interaction and privacy will be a point of contention pretty soon. At that moment, Qwen will be the knight in shining armor.
As someone who uses Opus daily for professional work and Qwen3.8-27B for all my private stuff, yes, Opus sub is faster for me, but I'm only rocking an R9700. If you're lucky enough to have sold a kidney on the blackmarket and purchased a 5090 and you're running ninfer, then actually... I think you'd be seeing fairly comparable performance!
I have actual obscure RGB corner floor lamps with BLE in my office. I had Claude find the BLE protocol and use my Office 365 calendar feed to change their color to red when I'm in a scheduled meeting (event), and back to warm white when the meeting is over, so co-workers know I'm busy without doing "are you busy" sign language through my window. The script is polled every 60 seconds. Tinkering with stuff like this is a lot of fun.
We've added some things like CI, Windows support, and starting some work on path-based permissioning for monorepos. These are still in the early stages so not documented that great yet and need to be tested a little bit more. Should have more updates in the next couple weeks.
They then give what they call a "hypothetical example but exemplary" of messages encoded in URL paths on a shared index page: "agent-07: answer(Q12)=42; need answer(Q19)=?".
So that's a GET request being used to pass information back and forth across multiple rounds. That's basically the DSEWiki pattern exactly. They say this likely came from the agents generalizing what they had learned from training with the official multi agent collaboration tool.
The report called it "misaligned behavior in an outlier scenario involving a rare and unexpected confluence of events."
They never mentioned a wiki, but this is most certainly it.
reply