Hacker Newsnew | past | comments | ask | show | jobs | submit | mlmonkey's commentslogin

<wrong thread>

Sorry, I think I'm being dumb, but where is the "first principals thinking" in your example?

They don't even have to "raise the prices back up". They just need to ensure that the number of output tokens goes up needlessly. Just make the model more talkative (via RL?) and cha-ching!

Oh yeah totally, good point. I remember there was controversy around Claude tokenizer generating more tokens lol

What about Jev? :-D

You can say that again!

This is the sad truth.

A lot of people are commenting about "externalities". That's just handwaving.

The real solution is to increase the ridership and the frequency. Public transit must be _reliable_. I live in SF and experience Muni's unreliability first hand.

The reason a person would prefer to drive over taking the Muni is a simple one: she knows the car will be there when she's ready to head back. Muni, on the other hand: one never knows when the next bus will come or if it will come at all. And then the City hands out Muni passes to all homeless people, regardless of behavior. Yesterday I sat down in a Muni train and guess what? A used piece of TP (about a foot long) was blowing around on the floor and everybody was playing dodge-the-TP, trying to jump out of its way. Buses and trains are often not cleaned at the end of the day.

Question is: why doesn't the City subsidize Uber/Waymo etc. also to the tune of $6.37/trip (inside SF)?


> What empty, bare land?

Have you been to Death Valley, my friend? All of Southeastern California is empty, barren land. Death Valley, Mojave Desert, etc.


> Have you been to Death Valley, my friend?

Well, no, but I’ve been close and I can imagine some features of the area that might have impacts on cost of construction and ongoing maintenance that are related to why it was empty land even before it was a National Park (and I can think of additional problems with California trying to build solar generation in a national park—different ones under the current federal administration than would be historically normal, but either way it would problematic.)


If you look at the city of Mojave, Calif. on Google Maps (satellite view) you can see massive solar farms all around it. There is plenty of space for expansion. There is no need to use a national or state park. (To be clear: Mojave Desert is far away from the city of Mojave!)

Guess how fast the wind blows across the surfaces. There's a reason why these places currently host massive wind turbine farms.

The wind and blown sand is something to behold. But large stretches are getting solar panels, yes.

They should give credit where credit is due. Solar panels on top of irrigation canals was pioneered in the Indian state of Gujarat back in 2012[1]

[1] https://www.autonocion.com/us/arizona-solar-panels-irrigatio...


They don't mention Gujarat specifically, but they do mention the prior projects in India and Arizona (which your link discusses) in the article.

You need to credit someone for "hey lets use solar panels to shade things that would benefit from shade"?

The article starts off by claiming that this person came up with the idea out of the blue; when in all likelihood they probably read a newspaper article about this being done in India. I remember reading about it then, which is why I was able to look it up and find the article.

It's just good manners to give credit where it's due.


What's wrong with mentioning it?

I noticed Terry Tao is not in this "advisory group" ...

Given that he apparently feels exploited after an appearance in some of their earlier marketing material, I imagine he's not overly keen on working more for their PR department, or this extension thereof.

> During this event, OpenAI requested an interview concerning my vision of the future of AI and mathematics. I accepted, and spoke with them for perhaps an hour. I had done similar interviews in various venues, and I assumed that, as with these other cases, they would eventually post the entire interview online, which talked about both the possibilities and risks of AI much as I have done in these other interviews. As it turned out, they only used a few snippets of that interview for that infamous advertisement instead.

https://terrytao.wordpress.com/2026/09/07/finite-time-blowup...


Why is the domain `pdp1173.com` when the machine is a PDP 11/83 ?

There's a couple eyebrow raising things about this post.. like the fact that there's nothing but vague details in text and one photo with an AI prompt like alt-text. I think they're just trying to get clicks for that banner ad.

Sadly, the site does seem AI-flavoured to a rather off-putting extent.

And, as you say, the details are so vague - it's exactly what you might expect if an LLM wrote them based on reading a spec sheet rather than any practical experience.

In particular, connecting that DEQNA to a modern network must have been a huge challenge all in itself and I'd have expected there to be a series of blog posts about just that. Maybe it's legit (and I'd love more detail if it is), but right now the presentation screams "thinly disguised ad".


I think he has actually posted videos of the restoration before (they've popped up in my youtube feed in the past), but I've never watched them, so dunno. The LLM site design and advertisement front-and-center are certainly a... choice....

But, more importantly, why do you reckon DEQNA would be hard to connect to a modern network? NetBSD still supports it when running on a q-bus VAX: https://github.com/NetBSD/src/blob/trunk/sys/dev/qbus/if_qe....

It has some pretty broken behavior regarding multicast/broadcast traffic IIRC, but nothing putting it behind a vlan or separate LAN altogether wouldn't solve. I've got a little lsi-11 system under my bed I'm hoping to, one day, get the Fuzzball [0] distribution running on. The 22-bit address extension wirewrap the previous owner did needs to be redone entirely, so it's been a "later" project for years now, but I am hoping my little DEQNA will be able to ping the Internet still.

[0] https://en.wikipedia.org/wiki/Fuzzball_router


I was assuming that something of that vintage would only have been able to handle a 10BASE5 transceiver, but a wee bit of googling suggests that an UTP transceiver works well enough.

Needs a patched ROM and recompiled kernel, though! See http://www.cosam.org/computers/dec/pdp11-23/20080402.html (and the two subsequent installments)

(That cosam.org site is more what I'd be expecting for a project like this, but maybe I'm just too old and grumpy for my own good!)


Dave posts frequently about his work, some on YouTube and some on Facebook. You have to be in the right groups on FB.

You and several other people here are making unsubstantiated claims about how you're super sure this is real in a way that makes it seem less real. You guys seem to know this Dave character and are doing him a solid, right?

Dave Plummer's a well known YouTuber. He pops up in a couple of DEC groups that I'm in on Facebook. However I'm not a fan. He seems to take a lot and contribute only self-promotional videos.

The photo on this page looks very AI heavy. I'm not an expert but I really don't think those light panels are real (if they're real they're not contemporary to the machines they're in).

(Edit: I should qualify that I don't mean the panel with the octet switches, that looks roughly right - I mean the one under the 8" disks and the one on the vaguely Vax-pedestal looking thing)


Okay gotcha, sorry I was jumping to conclusions.

Given the AI-heavy page, I don't think your skepticism was unreasonable.

While probably it's legit because I've seen Dave discussing the process on the YT channel (which I no longer watch) he unfortunately does have form https://news.ycombinator.com/item?id=39813625

Having TMOG on a .org when it sells commercial licences doesn't sit well with me given the history.


FWIW, .com was taken so I registered .org, and it's only used for the free version. The PRO version lives at tmog.store, which should be pretty plain.

> The PRO version lives at tmog.store

https://tmog.org/pro.html


Oh, yeah - this popped up on HN yesterday too and also met with scepticism: https://news.ycombinator.com/item?id=49772362

Oh well, so this PDP restoration probably isn't real. Disappointing.


"Probably isn't real". Interesting theory.

Well be real, you could do a lot more to make this convincing. Couple the lack of details with some other posts you've made that seem to be an attempt at advertising your software (the one with a big banner on this page), sprinkle in the AI generated website and the fact that the domain name says a different model than the page does.. this is going to raise some eyebrows.

You put all this work into a project like this, all you would need to do to prove it's real is post a selfie with the hardware.. but "interesting theory" is all you're going to say?


What's not real about it?

I'm a little confused by your comment.. especially considering the fact that you edited it and changed the link from and old comment thread from 2024 to whatever this dead post is from 2012. What does having form mean? Who is Dave?

Sorry, lost the last digit, fixed now. This restored PDP-11/83 project is Dave Plummer's

Because it started as a 15Mhz 11/73 but I later upgraded it to an 18MHz 11/83 CPU. And now it's a Mentec M1, but it's still an "83 class" machine. That's my thinking, anyway.

Anecdotally, I have found the same. I spend a lot of time with these frontier models, brainstorming, etc. and the drop in performance from, say, week 1 to week 8 is often massive. Whereas in the beginning, it seemed like a capable research assistant, by the end of week 8 or so it starts acting like a puppy dog eager to make its 'master' happy for a few treats.

Reminds me of how slot machine users swear the odds have changed on a machine.

also when someone says you just have to prompt it a certain way it reminds me of people who think they can get better results out of a slot machine by pressing buttons in a certain order

The providers of these models also design the UX similarly to slot machines (run it x amount of times for better results, multiplying your spend) this isnt a coincidence and they're playing into the gambler mentality, and probably hire UX designers that specialize in this.


I am skeptical of this as well but slot machines are programmable and the house can change the odds.

Well, running an LLM X amount of times does give you better results provided you are willing to select the best one out of the X yourself.

But I agree with your general point. One of the reasons subscription plans are cheaper because they modulate usage in this way based on demand. They can also recover compute more coarsely via usage resets (which give positive PR).


At that point I might as well do it myself

Well yeah. If for some task you find it easier to just do it yourself then you should. But you can improve the situation even if you can't entirely automate it by automating parts of the verification thereby making it easier to human-do larger verifications when X>1. But in many cases even that is not possible.

The progress however is such that the number of tasks that you can do with >p% automated and X=1 keeps increasing. So many times just waiting works. Of course, here also it changes from field to field. There are some tasks at which AI hasn't even gotten started, others where it has already peaked, others where it's increasing slowly, and others where it's increasing fast.


Wtf are you on my dude. Anthropic UI is designed like a slot machine? Hiring slot machine specialists? Sometimes I can’t believe im even on HN anymore with comments like this.

I think some of it comes from that they do not publicly let you see the random seed. So each time you ask the answer is different (like a slot machine) and if they let users use the random seed it would let people much more accurately assess if an underlying model changed somehow (same seed and same input will always have the same output).

Of course the closed Anthropic would never share this, it would definitely take away the 'magic' feeling of the AI


With models there are a bunch of other dials that can be tuned even if the model itself remains exactly the same.

Are those dials set the same across all hardware configurations and clusters? Does model behavior average out the same across different hardware?

There are just too many different buttons that can be set to really trust a provider either not to directly commit fraud, or indirectly commit fraud with system complexity affecting the output.


To my understanding, with batched inference and other "optimizations" you wouldn't get the exact same token predictions even with temp=0.0.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: