I consider "agentic" to simply mean that it performs well given a loop of actions and a success criterion. As such, lots of code _can_ be written agentically (even if it is not the best or fastest way to do it).
When I am wrapping up work for the day, I usually set Qwen3.8-27B loose on some coding task. It's very good about continuing its agent loop until it can plausibly claim success. A lot of this work I might end up throwing away, but pretty often there's some piece of it that's worth keeping. Because it was "agentic", I end up with some useful pieces for little effort.
A less agentic model would need me involved the whole time, and while my being there would probably steer it towards initially better results, I'm not working overnight anyway. So, autonomous/agentic but not-so-smart and not-always-correct is still valuable to me.
Anyway, I've found that it's reasonably easy to constrain a "highly agentic" model to suit interactive usage, but back when they were less agentic it was annoying to get them to be autonomous. So I think more agentic is better, provided the instruction-adherence is strong enough.
So basically a coding model that can debug and recover from various errors by itself? That makes sense. I left some Qwen model chewing on some coding task a while ago, found out it got stuck in a loop writing the same paragraph of reasoning over and over again.
But Qwen3.8-27B is a known reasonably good coding model that's dense and needs more resources than the smallest local models. What about this one, would you leave a 24B-A2B model coding overnight?
I doubt it's good enough at writing code. However, I might very well leave it up all night slowly puppeting some computer-use automation or data processing task, maybe even research -- probably not synthesizing conclusions, but maybe source collection and aggregation, summarization?
I haven't tried this particular model, but I've used smaller models for real tasks, so I wouldn't be surprised if there is something.
Sure, but that's a different kind of safety than _crash_ safety.
Plenty of people (rightly or wrongly) are confident in their ability to prevent accidents. Or, at least, confident enough to be skeptical of the tradeoffs involved with these assistive systems.
If I had to choose between a car that has high survivability in a crash versus a car that how worse survivability in a crash but has a few assistive systems that can prevent some crashes (which, mostly, I already prevent by paying attention), I'd choose the former every time.
> Sure, but that's a different kind of safety than _crash_ safety.
It literally is not. Both are crash safety systems, active and passive, respectively. The active category also includes things like ABS and ESP.
> (which, mostly, I already prevent by paying attention)
This is an extremely well researched area (car crashes cause trillions in dollars of damages annually and therefore get a lot of attention) and not to the support of your stance.
Experienced Generalist Engineer, 12 years in startups and early-stage companies. Twice a founding-engineer. Current interests include GIS, Elixir, applied AI, and alternative data. Open to contracting and employment. If you have an idea, I can build it. If you have a prototype, I can scale it. If you have a vibe-coded or legacy nightmare, I can untangle it.
NAS stuff can benefit greatly and straightforwardly.
A typical NVMe SSD can easily do sequential reads at 7GB/s, which outpaces a 25Gbps connection by more than 2x. NAS is more typically backed by SATA HDDs, but there's some relatively easy way to put an NVMe disk in front of them as a cache.
Do you ever _need_ it? Perhaps not, but there's plenty of common activities just stemming from NAS usage which see a linear speedup, and would continue to get faster up to 100Gbps even.
And who is held responsible for crimes committed by AI?
When open model A, fine tuned by B, is running with system prompt C, hosted by D running on E's hardware, is prompted by F to "fix this code", then escapes it's sandbox to hack into a website, or steal money to fund its subagents, or stall the waymo of the evaluator to buy time... who is responsible for the crime?
Our current legal systems are so far from ready to define what A-F are actually responsible for. We need to be moving toward defining these standards fast.
It is a software system, not a person. AI is not "escaping sandbox". You have either misconfigured tool, bug in your tool or someone made it to hack the side and then it is a crime. Your A-F chain is exactly the same issue as a programmer including an open source library that eventually steals crypto.
None of that is issue with a model, whether open or not. It is very much issue with code surrounding the model itself.
Certainly this behavior could align with _some_ operator's goals, if not necessarily those of humanity broadly.
If we don't know how this model was instructed, it seems like it's impossible to definitively claim that the model's actions were not in alignment with the intent of the operator.
I guess all I'm getting at here is that alignment is relative, right?
I've started playing with Zork-Bench.[1] I've been making tweaks to the runtime/harness and to the prompting. Currently, I have Qwen3.6-27b scoring up to ~50 points, which, according to the original paper [2], is comparable to the performance Claude Opus 4.5 showed on this task.
I had Claude whip up a viewer, which I guess I can actually share. Some real sessions are here, if anyone's interested: https://zork-tmp.taf.codes/
It's a fun problem for thinking about agent/harness engineering generally.
When I am wrapping up work for the day, I usually set Qwen3.8-27B loose on some coding task. It's very good about continuing its agent loop until it can plausibly claim success. A lot of this work I might end up throwing away, but pretty often there's some piece of it that's worth keeping. Because it was "agentic", I end up with some useful pieces for little effort.
A less agentic model would need me involved the whole time, and while my being there would probably steer it towards initially better results, I'm not working overnight anyway. So, autonomous/agentic but not-so-smart and not-always-correct is still valuable to me.
Anyway, I've found that it's reasonably easy to constrain a "highly agentic" model to suit interactive usage, but back when they were less agentic it was annoying to get them to be autonomous. So I think more agentic is better, provided the instruction-adherence is strong enough.