hephaes7us

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.

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.

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.

Although impractical, wouldn't that be nicer?

It looks like this is a bit less concerning if you have turned off open account registration.

Location: US, but flexible on TZ

Remote: Yes

Willing to relocate: No

Technologies: Infrastructure, Backend, AI-Applications, enough frontend for prototyping. ( Python, Elixir, Phoenix, Javascript, GNU/Linux, AWS, Postgres, DuckDB, etc )

Résumé/CV: https://resume.taf.codes/resume.pdf

Email: [email protected]

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.

I write at https://taf.codes.

Happy to discuss work for hire, and to connect with anyone having complementary capabilities.

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.

Crime is already illegal.

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?

Very cool. Do you happen to have any kind of mailing list, blog, etc?

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.

[1] https://github.com/mnky9800n/zork-bench [2] https://arxiv.org/abs/2602.15867

Location: US, but flexible on TZ

Remote: Yes

Willing to relocate: No

Technologies: Infrastructure, Backend, AI-Applications, enough frontend for prototyping. ( Python, Elixir, Phoenix, Javascript, GNU/Linux, AWS, Postgres, DuckDB, etc )

Résumé/CV: https://resume.taf.codes/resume.pdf

Email: [email protected]

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.

I write at https://taf.codes.

One thing I'm working on lately is semantic trend analysis across podcasts (using text embeddings). I just finished a custom NUMA-sharded vector scan kernel that can outperform DuckDB by 4x (on my specific queries). My audio transcription system is 10x cheaper than the most competitive provider, and I'm currently working to scale up my ingestion pipeline. If you're curious, my most recent post about this is here: https://taf.codes/2026/05/03/podcasts-2-trends.html

Happy to discuss work for hire, and to connect with anyone having complementary capabilities.

Farmers and pet-owners might prefer the rats.

Location: US, but flexible on TZ

Remote: Yes

Willing to relocate: No

Technologies: Infrastructure, Backend, AI-Applications, enough frontend for prototyping.

Résumé/CV: https://resume.taf.codes/resume.pdf

Email: [email protected]

Experienced Generalist Engineer, 12 years in startups and early-stage companies. Twice a founding-engineer. Current interests include Elixir, GIS, 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.

I write at https://taf.codes. Most recently I've been working on a trends analysis system for podcasts. I've transcribed hundreds of millions of words of podcast audio, embedded it, and can make breakdowns of topic prevalence by-time and across podcast feeds. I'm currently working on speeding up transcription with spot-GPUs, and speeding up analytical queries using a columnar data store. Initially that might use DuckDB, but I think I can do it significantly faster with a custom NUMA-sharded vector-scan kernel. Some discussion and examples here: https://taf.codes/2026/05/03/podcasts-2-trends.html

Happy to discuss work for hire, and to connect with anyone having complementary capabilities.

Same place it always does, just print it. Of course, that still effectively penalizes those who don't want children, but the penalty is less legible to the public so there are fewer objections.

I do too, but it does seem like fighting the tool. I wish more people would use email.

Isn't there a meaningful sense in which "separate VMs for different apps" constitutes a cluster?

The "cooperative task" they're engaged in is just, broadly, meeting your needs, whatever they are.

The isolation is a desirable property, and I agree this is much preferable to a highly inter-coupled bunch of machines, and also that thia stretches the typical sense in which we refer to a "compute cluster", but I don't think it's an entirely invalid framing of the term.

You can't, really. If a user can access the site, so can a bot.

You may be able to make it more expensive than your information is worth, but of course that affects users too.

If you don't see people talking about the "end of personal vehicles", it could just be that you haven't looked very hard.

It's intuitively obvious to a lot of people that the era of personal, wholly owned transportation is waning. A lot of people seem to miss the second clause of that old "you'll own nothing" phase, the part where most people are happy about it!

When vehicles drive themselves, and there's a large enough pool that one can show up pretty reliably within a few minutes of your needing one, how many people are going to choose to own when renting is cheaper and easier?

Oh, I suspect you can overpower it, if that's what you mean. However, if you weren't expecting it'd do something like that in the first place, it can accomplish a nonzero change to your trajectory before you intervene.

Believe me or don't, but, if you operate a vehicle with these assistive systems, I encourage you to carefully familiarize yourself with the ways in which they may unexpectedly affect its behaviour.

When the car actually drives itself completely, I think they will be safer than human drivers.

All of these half measures are pretty concerning to me. I think they let drivers feel more comfortable, despite paying less attention, and I think their failure modes may often be much worse than the (human-driven) crashes they purport to prevent.

Anecdote: I once had a rental car with alane-keeping assistance system that would nudge the wheel slightly. On the interstate, upon cresting a hill, I saw that there was a vehicle stopped in the shoulder, and I was concerned someone might step out into the travel lane. I already knew that there were no vehicles behind me in either lane, so I steered gently into the passing lane to give ample space to anybody who might step into the road.

However, in my haste, I had not used the blinker, so the lane-keeping system intervened. Imagine my surprise when the car decided to nudge me back towards exactly the dangerous situation I had been avoiding!

Luckily, nobody stepped out into the road. But if they had, this lane-keeping system could have killed them.

In comparison, even if the left lane hadn't been clear, the hypothetical accident there would have been a comparatively minor fender bender.

For years, vehicles have had a little light that comes on when you are below about 50 miles of range. It's next to the fuel gauge. I've always heard it called the "walk light", which I presume is a reference to the fact that, if you don't do something, you may have to start walking soon.

My car has a little screen in the dash where it usually shows my range, or the current temperature - information that I check when safe to do so, but never very urgently. This is the perfect place for a warning about low wiper fluid.

As for forward collision warnings, ehhhh. Maybe that should beep loudly, but it should almost never be wrong! (A false alarm could easily mean I slam on the brakes and get rear-ended, so that has to be balanced with the safety advantage of the true alarm.)

I can understand why it'd be preferable to avoid such a bridge layer, and indeed I too would rather just have a transparent view of what's going on at the protocol level.

Stability and performance at scale sound like implementation specific properties though. If you've tried this, I'd be curious to known about the specific issues you encountered.

Something like Rclone and a cron job, or else s3 mounted via FUSE, could possibly bridge that. Of course then you have to worry about reliability of the bridge...

What kind of cases were you measuring? I would think that, e.g. 256 separate long-lived connections in a setup like that would scale less-than-linearly but not dramatically so?

Though it may be painful for much of the world to move on from Microsoft, at some point it could be more painful for them to stay with Microsoft. The inertia is huge, but inertia doesn't carry anything forever.

You could probably serve 100k visitors from a $5 VPS, depending on the application.

That said, I understand people are paying for basically not having to think about infrastructure, and agree that that's theoretically worth money, if they could do it well.

I don't think the above poster is talking about finding novel treatments, but rather that they're talking about aiding in diagnosis and navigating existing treatment options.

We always wish that our doctors would stay up to date on all of the current medical literature as they practice, and some of them do. In theory, AI systems could greatly accelerate a person's ability to retrieve and extract insights from the current body of knowledge.

Of course, that is highly fraught, but, in theory, I think I see what they're going for.

Speaking in terms of GW makes sense when we are discussing performance of an individual datacenter since power is, kind of, the limiting factor.

Also, FLOPs per watt hasn't changed as much lately, so thinking in terms of watts over a few-year time horizon does at least give you a ballpark of how many FLOPs.

If you're talking about a chip, you obviously want to know about FLOPs moreso, but, even down to the level of individual rack-units, wattage is a serious concern. Not every facility is built for these crazy 200kW racks.

This is essentially a smarter auto-clicker. I'm not sure I'd call it "reverse engineering".

A TOS/CSA should in no way ever attempt to prohibit automation, and if it does, it (generally) deserves to be disrespected.

There is a legitimate concern however about customer resource use escalating beyond what was expected when the price was set. Luckily this can be written as a simple black and white determination without any complicated gray areas, and is therefore easily enforced both in the code and in the contract.