As an experiment I took, not this current quant, of Qwen 3.8 27B and had it analyze some of my writing in CPU-only mode with llamacpp. Tried a Q4 and Q5. It was dreadfully slow, but you fire that off before bed, and you can get results the next day. This was a ~6000 word markdown file, took maybe 10 minutes before it started generating anything at all, but then it just cranks away. I would need to tune down the default thinking intensity, but it otherwise did the job.

Not compatible with vibe coding, but if you want a local tool to review content, it is possible today.

FWIW depending on what exactly you want to review something smaller like Gemma 4 E4B would be much faster. On my 7yo 3700X, llama.cpp runs Gemma 4 E4B UD-Q4_K_XL with a 131k context (context size doesn't seem to make much of a difference in speed) at ~10-15 tps (using MTP draft).

I'm mainly using it when i need some simple stuff and Qwen 3.8 27B is already occupied in the background :-P

I kind of like the slow process, it forces me to be deliberate and not lean on the tooling as a crutch. I don't write that much, so having to wait a day is kind of like having your final work ready to review by a human editor.

Plus, using a bigger model makes me feel a little better that I am not taking a hit on quality when I do not even care about the turnaround time.

I’ve been using Bonsai 2 (also based off of Qwen3.8 27B) on my mac mini in this way (using pizza-bot for the harness). In my case I’ve been using more for deep web search type stuff. It’s been a nice workflow to kick off a request in the evening and come back to a 1 page report on it the next morning.

It’s fun, but it’s just so tough to see the value when the same research report via ChatGPT is done in 5 minutes and likely higher quality, for less than a dollar a day.

This strikes me as evolving resistance to fire. Sure, it would be a great adaptation, but there is only so much a body can do to mitigate a concentrated burst of energy.

More likely is potentially better camouflage or locating on an area of the fish which is harder to target.

  But Esvelt points out the policy only requires screening for DNA sequences 50 base pairs or longer. A would-be bioterrorist, he argues, could easily stitch together a dangerous sequence from much smaller fragments ordered from multiple different suppliers. Esvelt says he has built an algorithm capable of screening sequences as short as 30 base pairs for their risk potential and made it freely available. “That’s the main flaw in an otherwise pretty great policy,” he says. “Don’t bother building a fence if there’s going to be an obvious huge gap in it.”

So, a song and dance of doing something that will most likely just inconvenience people doing real work.

Also, DNA chemistry is very well established. Someone with the appropriate background could easily assemble oligos without too much difficulty. I guess this stops Little Johnny from printing ebola from the internet.

>So, a song and dance of doing something that will most likely just inconvenience people doing real work.

Why does he call it an "otherwise pretty great policy"?

He's claiming that the policy needs to be more comprehensive. You're concluding that it doesn't pass a cost/benefit test. That's not at all the same claim!

Classifying into one of 20 groups? Did anyone try scikit learn?

I did this at my last job. ComplementNB trained on ~30,000 examples (which were freeform text, and very unbalanced across categories) in under a minute on a laptop CPU, and with log-weighting of scores, was about 85% accurate. I don't remember how fast it was at evaluating the tickets beyond 'basically instant', as it was a proof-of-concept that got cancelled by changing ticketing system.

Heard about classifiers a long time back. Will current ones resolve "reallyyyyyy beast mode" as a positive phrase like an LLM?

I was going to say, and I say this a little bit too often, a BERT model can do this even faster, with even less memory...

The normie use cases need less performance, so it could be fine. Review this email, presentation slide, make a flyer for the soccer game, what should I do in Barcelona, etc. If you are not running gas town, you don't necessarily need that much.

That narrow desk makes me uneasy. Seems like a mini ordeal where you have to perfectly line up the chair to insert and extricate yourself.

Those guys were fast on the draw to cover the body with those sheets.

He’s hired to follow the robot and put the sheet on it whenever it’s dropped down to the floor

Now I have a weekend project in the future, assuming I can find a compatible process for my hardware. Never connected to the network, but the thing makes a few ever annoying popups I would love to disable. Or at least ensure that no "helpful" guest fills in the wifi information when I am not looking.

That scene from Elf where Will Ferrel discovers the World's Best Coffee.

It is impossible to avoid any impact from advertisements, but the population has become somewhat* inoculated after a lifetime of exposure.

*Minus the X% of people who are sold on the most recent BS they have recently heard.

Millions of people on TikTok and the lines after a place has been hit by an influencer implies that this inoculation is wearing out.

But that works by hiding an ad in a context that not everyone realises is an ad, bypassing the conditioning.

We used to have a turbopump that would trip the breakers every time you turned it on from cold. The pump seemingly attempted to spin up to max RPM before settling down to a more sustainable load. The only way to get the thing to stay on was to turn it on once, let it warm itself up on the initial attempt, trip the breakers, and try it another cycle now that it had some momentum and would pull slightly fewer amps. Sometimes took a few attempts to get that rolling start working.

That's a great way to lose the breaker though. If you're lucky the latch will fail. If you're unlucky it will weld shut.

Not ideal, but it was the only thing that worked. This was attached to a million dollar piece of equipment, so any remedial electrical replacement work would have been slightly inconsequential. Thankfully, it was one of those machines that you turned off as infrequently as possible, so we did not have to do this procedure too many times.

The vendor did get the hint, and the second iteration of the device did slowly warm up the pumps so this process would no longer be required.

Yes, soft start is the way to go. Also: proper industrial breakers allow for temporary overload. I have a couple of pieces of gear that have inrush currents well over the limit of a normal breaker so I put them on a slow one and now it is fine.

https://onccy.com/how-to-choose-between-type-a-b-c-and-d-min...

> Also: proper industrial breakers allow for temporary overload

Notoriously a crescent wrench makes an excellent 1000A slow blow fuse

A 5/32" flat-blade screwdriver makes an excellent fast-blow fuse when dropped across exposed bus bars. And you can re-use the remains of the handle to stir your coffee once you've retrieved it from across the room.

Styropyro shows something like that!

https://youtu.be/OC7sNfNuTNU

Soft-start is in theory the way to go, but it's very dependent on the motor type, there's a whole pile of different methods required for induction motors depending on whether they're capacitor-start, permanent split capacitor, shaded-pole, etc, so you can do it in some cases by lowering the voltage, or with a VFD, or a VSD, or the old-fashioned way with a centrifugal switch but that requires very careful handling. If as the OP said it's running million-dollar equipment you probably don't want to monkey with it at that level.

Reminds me of computer fans, some bioses seem to do that ramp up for no reason thing as well.

Fans have a minimum speed needed to spin up when cold starting from 0 RPM. So it's safer to start at near max RPM and spin down to the desired target RPM if the controller doesn't know the exact operating characteristics of each fan.

You can observe this with a speed control knob/ESP32 where a running fan will happily run forever with the knob turned down to e.g. a 30% duty cycle. But after turning to 0% and waiting for the blades to fully stop, if set back to the same 30% they won't move at all or you'll maybe hear a sketchy clicking/whine sound with the blades intermittently twitching.

I've always assumed it was to calibrate the PWM sensor for fan speed, so it starts with no sensing active, runs up to max, and then the sensor kicks in and settles it down. You don't need to run a fan at max speed to get it to start, most will start with very little power applied.

not exactly relevant to computer fans, but in case of your usual standing fan you need a way larger kick to get it going. if it fails to start but stays powered it will stay stuck, heat up and fail. this is why fans usually have the highest option as the first one, so you have to start it at full speed at first, and only then reduce it.

You also want to be sure that, in the absense of fan-control, the CPU won't overheat. Full-on is safer than nothing until a driver is loaded.

Isn't that what a soft starter relay is for?