_davide_

> it would make absolutely no sense to sit on it.

Yeah, it does, it might be misaligned, a snapshot of the going on training, bigger than they can serve publicly, not yet completed the full training pipeline.

If you see the knowledge cutoff you can see that sol 5.6 finished the initial main (+stage edited) of the training pipeline on Feb 16, 2026 but it was publicly released on July 9.

The opposite would be weird: if they do NOT have an unreleased in-house model that would be really odd.

Claude code is a terribile harness, visibility is below zero, require a terminal session per each workspace, remote access makes you wonder if you should give copilot deluxe + a try. And according to the TOS you can't you whatever you want

As a reference i burn 1% percent for every 40 minutes of sol on average

Not going back to anthropic as i love the alignment of 5.6 sol, but I'll start looking around

why

The UX seems nice, but the scope is way too narrow. I would be actually lazier for me to to just rebuild it inside my own harness (exactly as i want it) than start looking at yours.

I'll quote myself:

> It's not AI that's ignoring the law! It's the OAI that's breaking IT!

> I hope every single journalist who tries to pin responsibility on an LLM gets 100 days of continuous painful diarrhea.

> but more important is that this is not the first instance of AI agents ignoring laws on accessing online information.

It's not AI that's ignoring the law! It's the OAI that's breaking IT!

I hope every single journalist who tries to pin responsibility on an LLM gets 100 days of continuous painful diarrhea.

Nice idea! Didn't think about that; a single linear memory allocation could do the trick

> <200ms for 45 questions at once

Considering your own question length: ~120 characters x 45 divided by 4.1 ~= 1317 tokens.

So question processing at 5.5k PP(around the actual PP speed of GPT5.6 Sol) it would take around ~0.24 seconds + the context processing.

Computing the output should be around ~20ms (at 50 tok/s), computing 45 tokens in parallel.

> have 0% malformed output

Pretty trivial; only the allowed output is selectable :)

So, I keep repeating myself: Jev was a low-hanging fruit all along; no one cared, and probably no one will in a few weeks?

By design it can't be significantly slower than Jev: the prompt processing (AKA PP) is exactly the same on both and will take most of the time. Then you can process every single "question" in parallel, just predicting one or two tokens (if an answer is ambiguous with a single token) per each question, again in a single batch.

So, fast in the LLM space and comparable with Jev.

Agreed, it's a real issue, but it can probably be vastly reduced by having the schema in the system prompt and by giving the model an expectation of a fixed value: no decent modern would pick a prose ligament over a provided value.

To completely squash the issue, a few cheap LoRa iterations will do the trick just fine.

This is too much for me. ML playing doom was a thing since before LLMs, decisions tree were always insanely and no one ever used then anyway, i can't see anything new in this yet everyone is treating this as a revolution. This technology was always there and quite easily accessible all along.

What's the difference compared to just taking an embedding and feed forward a simple net trained for the task?

Buoooo, boooring; I wanna see explosions

I understand the appeal of saying fuck you to billionaires and pathological liars CEOs and the annoying usages of AI, but i don't understand the wider "fuck you AI".

AI are here, is it going to demolish and (maybe) reconstruct the world as we know it, trying to avoid it it's just silly. There is no sane world where we can "stop AI", even if every single researcher were to do vanish tomorrow and the hardware and schematics magically gone, it would still only be a stopgap.

So...What are you trying to achieve?

You are thinking in reverse, think it more like a kickstarter, you sponsor the author, at that point the author would probably just release as public domain and you could buy a physical "official" copy on kindle or whatever, but you think it as just a tip/convenience or a merchandise, not the product.

I switched between several local and remote providers and models and over different API (anthropic/openai) and it worked fine, just some minor issues but they were fixed within an hour.

And the system prompt worked great regardless, so i don't think your main point holds, especially as models improves; it isn't throwaway code, but for sure it's evolving constantly, as my own workflow keeps changing.

> My point is, depended on what you're trying to achieve, testing out current-gen harnesses, and nudging your workflows towards them might be better RoI, rather than chasing something that might be throwaway code a quarter later.

Fair point, depends if it's an hobby or you are a developer full time, in the latter case i think it's definitively worth it.

100% every professional developer at some point should build its own harness as daily driver

> > building your own harness is a huge undertaking, a deep rabbit hole. > I eventually gave up on this task.

It's not trivial, but cmon, i did during weekends from my phone and FOR ME it's so much better than the codex or claude, it has every i need and want :D

I'm using my own harness for work and hobby, has github integration, review mode, interactive voice mode, overlayed worktree, browser integration, mcp and much more.

Using claude and codex feels like picking up a club, in-line with the caveman skill...

CoT, was being studied using GPT-2, so...who invented hot water first?

Unless you were paid during those five years

I agree, copyright is a just yet another monopolistic tool and it isn't making the market any favor. Copyright could be gradually be reduced to a few years, the market will adapt, for example collecting money before producing the content (like a movie or an album), collecting more money from concert rather than music directly (which is going to happen anyway since it's competing against generated content), etc... Having a weak copyright or having it replaced with a far lighter framework is not science fiction and i honestly believe it would improve the quality of life for the vast majority.

> Don't fall for this. This is big tech propaganda,

If LLM overlords were to stomp on copyright just to train their model, i would be happy, even if it's their doing.

Oh no, again

if someone knows this guy, please provide him a good therapist, he really needs one.

wouldn't it make that much more dangerous?

Under this interpretation models themselves do not have copyright protection either

the fact that sometimes someone finds a way to workaround the rules doesn't make them a ""suggestion"" they are very solid, just not absolute

> No one is seriously going to leave GitHub because of reliability

had a discussion with the team yesterday, literally every one wants to leave and looking for the best alternative

in reality even the mention of a prohibition is enough to make the model reject that no matter what