My F-You To Sam Altman — Transcript
Full transcript
- 0:00Cloud AI is becoming an incredibly
- 0:02powerful tool. With the latest Frontier
- 0:04models, you can write shoddy documents,
- 0:06generate shoddy pictures, and even
- 0:08compose shoddy music.
- 0:11[music]
- 0:14>> Okay, jokes aside though, it can be
- 0:17legitimately useful for everything from
- 0:19writing code to data analysis to just
- 0:22asking it questions that are too dumb to
- 0:24ask your friends and colleagues.
- 0:25Personally, I've been doing a lot of
- 0:27that lately around smart home
- 0:28automation. And yeah, it's the kind of
- 0:30thing where I could read a 100 pages of
- 0:33documentation or I could just ask in
- 0:35plain English, hey, what does this
- 0:37parameter do? And even if it gives me
- 0:39the wrong answer, which it sometimes
- 0:41does, it usually provides useful clues
- 0:43that will lead me to the right answer.
- 0:45There's just one small problem. Well,
- 0:47actually many problems. hallucinations,
- 0:50privacy, the mass theft of the entire
- 0:52output of humanity, uh cost, and while
- 0:55running an LLM locally can't solve all
- 0:58of those issues, it can certainly solve
- 1:00some of them. By buying the hardware
- 1:02upfront, I'm not going to be able to
- 1:04avoid all token costs. Some things
- 1:06really do benefit from Frontier models,
- 1:08and it looks like that's going to remain
- 1:09true for some time yet. But I can
- 1:12probably do over 90 plus% of what I need
- 1:15to do on my hardware with no
- 1:18subscription, which has many benefits.
- 1:21The biggest one being that I can tie
- 1:23constant AI compute into other systems
- 1:26that I have without worrying about token
- 1:28cost overrun. And in doing so, I keep
- 1:30everything within my network where it's
- 1:33private [music] and will always work
- 1:35even if my internet is down. All I need
- 1:38to do now is build it. Um, what's a 5090
- 1:42worth these days?
- 1:44Oh,
- 1:47well, this is going to hurt.
- 1:49Just like this painful segue to our
- 1:52sponsor,
- 1:53>> Squarespace. They make it easy for you
- 1:55to build the ideal website to go with
- 1:57your brand or business. Create something
- 1:59custom with their design intelligence
- 2:01tool with our link below.
- 2:12As you can see, I already got the
- 2:13hardware for my home AI station mostly
- 2:16set up on my own time, including making
- 2:18some fixes to the uh Frankenstein power
- 2:21connector GPU here, that, by the way,
- 2:23worked out fine
- 2:26after several attempts. This is the
- 2:29worst thing we've ever done. So, the
- 2:30main reason I'm rebuilding it is
- 2:32actually because Silverstone reached out
- 2:34telling me that they have finally built
- 2:36an update to my favorite case of all
- 2:39time, the Tempen TJ07. And I was mostly
- 2:43just really excited to check it out. So,
- 2:45I'm going to put that in here. This is
- 2:47my little portable water cooling unit
- 2:48that I use anytime I need to
- 2:50troubleshoot one of like the rack
- 2:51mounted water cooling units if I have to
- 2:53pull it out of the rack. Uh, it's been
- 2:55permanently attached to the AI station
- 2:56lately, so that's been kind of
- 2:57inconvenient. On the subject of
- 2:59inconvenience, I noticed that all the
- 3:01tables in here are like Lucas height. Oh
- 3:04god,
- 3:06you know your case is [music] off the
- 3:08rails when you have these like TV style
- 3:10box closers. I used to do all the super
- 3:13high-end custom builds back at NCX in
- 3:15the predecessor to this case because it
- 3:17had the best cooling, because it was the
- 3:20easiest to build in. Um, no. And no, but
- 3:24because it was absolutely sick.
- 3:34Oh,
- 3:39and the new one looks glorious. Good
- 3:43lord,
- 3:45they wrapped it around even more.
- 3:51Oh, okay. Goodbye. Wait, what is that?
- 3:54Oh god, it's got wheels.
- 4:00Okay. One solid piece of aluminum all
- 4:05the way
- 4:08around.
- 4:10Is that sick or what? Only this flat
- 4:14back plate is a separate piece. Look at
- 4:18all that room for activities. And this
- 4:20is a freaking awesome looking GPU
- 4:23anti-segmount. As you can see, it's
- 4:26clearly designed with multi-GPUs in
- 4:28mind. In fact, that's one of the only
- 4:30reasons that Silverstone was able to
- 4:31bring this thing to market because
- 4:33[music] gamers don't really need
- 4:35multi-GPUs and fancy expensive tower
- 4:37cases anymore. But AI workstations, they
- 4:41do. So, it is packed with support for
- 4:44cooling as well. Where there used to be
- 4:465 and 1/4 in bays in the front, now
- 4:48we've got a radiator mount to go along
- 4:50with the radiator mount up here to go
- 4:51along with the radiator mount down here.
- 4:53So, we're going to be water cooling this
- 4:55today. And just like the original, we've
- 4:58got support for multiple power supplies
- 5:00for extremely power hungry builds.
- 5:01Although, it's worth noting we're not
- 5:03going to need that today because
- 5:04Silverstone also sent over their new
- 5:073,000 W 80 plus platinum power supply.
- 5:09[music]
- 5:10250 amps on the 12vt rail.
- 5:14That's stupid. This thing is
- 5:18It's like $1,300 or something like that.
- 5:21and it looks and feels every penny of
- 5:23it. As for the rest of the hardware I'm
- 5:25using, well, it was mostly dictated by
- 5:28whatever labs didn't take.
- 5:32The good news is the leftovers here are
- 5:34still pretty freaking amazing. Can you
- 5:37believe that absolutely nobody wanted
- 5:40this Sapphire Rapids workstation? It was
- 5:42just sitting on the shelf. I don't even
- 5:44know if we ever did a video on it. For
- 5:46the CPU, then I ended up with a Zeon W9
- 5:483495X. It's a 56 core processor that can
- 5:52turbo up to 4.8 GHz, which is nice, but
- 5:56not actually the reason that it's most
- 5:58suitable for a system like this. What
- 6:00mattered more to me is that it supports
- 6:02boatloads of system memory, which is
- 6:03perfect if you're running a ton of
- 6:05containers, and it supports a ton of
- 6:08PCIe Gen 5 lanes, 112 to be precise.
- 6:12Meaning that on a motherboard like this
- 6:14W790E Sage SE from ASUS, these are not
- 6:18decorative. These are almost all running
- 6:20at full by6. Not that I have that many
- 6:23GPUs to put it in anyway, but it sure
- 6:26doesn't hurt. [snorts] On the subject of
- 6:28GPUs, my goal was to find the ones that
- 6:30would be the most performant for me, but
- 6:32the least missed by the rest of the
- 6:34team. So, I grabbed the 5090 from the
- 6:37Murderbox PC because at the time I
- 6:40borrowed it, labs hadn't really delved
- 6:42into desktop liquid cooling, but now I
- 6:44can see they actually do a fair bit of
- 6:45it. So, um, glad I got this, uh, before
- 6:48they were using it.
- 6:53And then, because the goal was maximum
- 6:54performance, I went sniffing around to
- 6:56see if there were any more that I could
- 6:57borrow. And I found this. You guys might
- 7:01remember the video where we fixed
- 7:03Nvidia's much maligned 12volt 2x6 power
- 7:05connector by putting an XT120 on it.
- 7:09We did a pretty budgy job of everything
- 7:12from the soldering to our choice of
- 7:14wiring. And this isn't much better, but
- 7:16I fixed it up a little bit at home.
- 7:20Lucas went like this.
- 7:22I don't know what that means.
- 7:24>> I endorsed the tinkering. Let's say
- 7:26we'll see if it lasts long term.
- 7:28[laughter]
- 7:28>> It's been working great for now. With
- 7:31the GPUs out of the way, the last big
- 7:33question was RAM. Now, to run LLM, I
- 7:36don't strictly [music] speaking need a
- 7:38ton of system memory, but if I want to
- 7:40run a whole bunch of containers, then it
- 7:43certainly doesn't hurt. And this
- 7:45[clears throat] platform uses RDMs,
- 7:48which right now are pretty expensive.
- 7:51What the? Fortunately, I came across the
- 7:53kit that you guys saw installed when I
- 7:55was fishing around here late at night,
- 7:57and I went all the way up to this top
- 8:00shelf, which has
- 8:03lots of weird stuff in it, including a
- 8:06couple of boxes that were marked,
- 8:07[music] "Do not take without asking the
- 8:10business team." So, I took them. Intel's
- 8:14LGA4677,
- 8:16to put it lightly, was not popular with
- 8:18enthusiasts. So, going into this build,
- 8:21I had actually never mounted one of
- 8:23these before. And I was shocked,
- 8:26shocked, I say, to discover that instead
- 8:28of putting the CPU in the socket,
- 8:31thermal compound goes on the CPU. Then
- 8:33the CPU actually clips onto the cooler
- 8:40like that.
- 8:43Okay, then as wrong and horrible as this
- 8:46feels,
- 8:48the whole thing just goes on like that.
- 8:52Terrifying.
- 8:53Now, there's a torque spec for exactly
- 8:55how hard you're supposed to do it down.
- 8:57I believe we're going to tighten it
- 8:58incrementally in a cross pattern. There
- 9:01are some server platforms that actually
- 9:03call for you to just tighten them all
- 9:04the way down in a cross pattern. Not
- 9:07this one. And the spec is 8 inb for
- 9:09those of you who are planning something
- 9:11similar at home. All right. All right.
- 9:12Now, we need an airflow plan. I'm
- 9:14thinking exhaust out the top. Now, some
- 9:16of you are probably wondering, Lionus,
- 9:18what's the deal? Why are you not like
- 9:19water cooling this thing? And the answer
- 9:21is that not a lot of water blocks for
- 9:24this platform. Also, I don't really need
- 9:26to. I'm not overclocking. And these
- 9:28workstation chips do not run hotter than
- 9:30this Noctua cooler can handle. And this
- 9:32is a functional machine. It's not to
- 9:34show off.
- 9:38So, we're just going to Oh, figure out
- 9:41uh where all the hardware is for this
- 9:43case that we actually have a new
- 9:44motherboard standoff. Some moment,
- 9:45please.
- 9:46Why would they do this? They
- 9:50hide it all in the very top of the case,
- 9:52which if you remember is kind of the
- 9:55bottom of the packaging once you
- 9:56actually pull the whole thing apart.
- 10:00Is that it?
- 10:02Okay. Oh, wow.
- 10:05That's it. No manual, no nothing. I
- 10:09guess the vast majority of what you
- 10:10would install in here would come with
- 10:11its own stuff, though. Anyway, we've got
- 10:13one of those fancy wy little retaining
- 10:16motherboard standoffs. So, we'll go
- 10:17ahead and put that in the middle. I do
- 10:20not know the bolt pattern for this
- 10:21motherboard.
- 10:23It is unusual.
- 10:26[clears throat]
- 10:26Here's a tech tip I learned from my
- 10:28stepdad. Yeah, if you ever need to
- 10:29measure something, you don't have a
- 10:30measuring tape handy.
- 10:36That one.
- 10:41Now it feels like No, it feels like I'm
- 10:43up against something.
- 10:46Wait,
- 10:50how thick is this armor?
- 10:54Is that interfering?
- 10:56Cuz like I can't get it all the way
- 10:57down.
- 10:59Oh,
- 11:01see this is what happens when you're
- 11:03operating at the edges of standards. See
- 11:06these cable management grommets?
- 11:09They thick.
- 11:11See that?
- 11:13See this armor on the back of this
- 11:15motherboard?
- 11:16It also thick. So, we've got a bit of an
- 11:19interference issue. Fortunately, I see
- 11:22something. They're actually a little bit
- 11:24thinner on one side. So, what I'm going
- 11:26to try and do is turn these around and
- 11:29see if that resolves the clearance
- 11:30issue. Hey, there we go. Okay, that was
- 11:34much more successful this time.
- 11:38Can actually see the issue better from
- 11:40the back where the board is right there.
- 11:44That's the AR. Oh,
- 11:47well, ain't nobody using those cable
- 11:49management grommets anyway then.
- 11:53>> [laughter]
- 11:54>> A
- 11:57how has it come up so much recently? Oh,
- 12:01wait. I don't need it. Oh, Silverstone.
- 12:05Mwah. Beautiful. Amazing. They've got
- 12:09one of those pass through things, but
- 12:10it's wide enough to accommodate the LT
- 12:12screwdriver. How much you want to bet?
- 12:14Tony made sure of that. Love it. Thank
- 12:17you, Tony. And it goes just like that.
- 12:21That seems like enough cooling, right?
- 12:23We got an intake fan right there.
- 12:26[snorts]
- 12:28Man, this is [laughter] the kind of
- 12:30hardware that just like Okay, I don't
- 12:32have pit stains yet, but I will. I
- 12:34[laughter] just like seriously I think I
- 12:38was a couple WAN shows ago old when I
- 12:41discovered how much these are going for
- 12:43right now. Like obviously [music] one
- 12:45with this manner of modifications is not
- 12:47going to be worth as much, but like
- 12:51it's insane. And the craziest [music]
- 12:54part is that 5090s are not even
- 12:57necessarily the best choice depending on
- 12:59what you're doing. The GPU is obviously
- 13:02really fast and supports, you know, all
- 13:04of the latest types of quantization
- 13:06[music]
- 13:06and everything, but what they don't have
- 13:09in the grand scheme of things is
- 13:11actually super fast memory, nor do they
- 13:13have a ton of it. [music] So, what I
- 13:15might want to do in the longer term is
- 13:17actually like pull one of these 5090s
- 13:20and then augment a single fast GPU with,
- 13:24you know, maybe one or a couple of those
- 13:26memory unlocked crypto mining cards to,
- 13:29you know, run a gentic task or to run
- 13:32multiple models where speed is not as
- 13:34important. On that note, the labs team
- 13:36finally got their hands on some DGX
- 13:39Sparks. And this is a great example of
- 13:41how having the fastest processor is not
- 13:44necessarily as important as having tons
- 13:46and tons of memory that can be pulled.
- 13:48What are you guys using these for?
- 13:50>> Uh, today making cookie cutters.
- 13:54Seriously? Yeah.
- 13:56>> Okay.
- 13:58Right now it's a DeepSeek 4.1 flash on
- 14:00there which is like 410 GB of like
- 14:03parameters I think and then more
- 14:04context. So it's something you can only
- 14:06fit on there and it's quite fast and
- 14:09capable. So
- 14:09>> that's pretty cool. I tracked down the
- 14:12one 420 mm radiator in our entire
- 14:15warehouse. This old EK cool stream and
- 14:18ah
- 14:20got my pump and reservoir mounted.
- 14:22Silverstone has a really nice little
- 14:24system for this where there's just this
- 14:25plate at the front that has a couple of
- 14:28120 mm wide spaced holes. So, you could
- 14:31just throw a bracket on there and you're
- 14:33good to go. The one thing that I would
- 14:35like to see improved is there's nothing
- 14:37in the way of antiibration and on
- 14:39something that's clearly meant to have
- 14:40pumps mounted to it.
- 14:43Yeah, I'd like to see that like dampened
- 14:45in some way. probably going to be
- 14:47controversial, but yeah, we're going
- 14:50softline tubing for this. ZMT to be
- 14:52specific. It has really low evaporation
- 14:55through the tubing, and it's not a
- 14:58showcase build. It doesn't need to look
- 15:00like anything special.
- 15:02In goes our GPU now, or at least primary
- 15:05main water cooled GPU. Oh man, it's
- 15:08heavy. Oh. Oh, yeah. I might actually
- 15:13keep these fittings. These are my
- 15:15ancient Bits Power true silver fittings.
- 15:17And the main sticktick for these is that
- 15:20they're actually plated in silver. And
- 15:22silver has biocidal properties, which is
- 15:24great for maintaining a clean loop.
- 15:26These are old school. These are like
- 15:2820-year-old fittings or something like
- 15:29that. But that's one of the things about
- 15:31water cooling is that if you invest in
- 15:32really nice stuff, as long as you're not
- 15:33constantly chasing the latest fashions,
- 15:35you can reuse it for many, many years.
- 15:37And crazily enough, for not really any
- 15:40more than like a pre- AIO cooled 5090,
- 15:44we've got kind of like an AIO cooled
- 15:475090 for such a high-end board, there's
- 15:49really not that much in the way of front
- 15:51panel tom foolery. No RGB support even.
- 15:54So, this is just going to be a clean
- 15:56build.
- 15:58Oh, rats. I went and did the thing. I
- 16:01installed the GPU before hooking up
- 16:03everything and now I got to pop it out.
- 16:05It's okay. That's no big deal. Thanks to
- 16:07my trusty LTT socks. You thought I was
- 16:10going to say screwdriver, didn't you?
- 16:13Bam. No sweat. Or well, like you can
- 16:15have a bit of sweat, but like you know,
- 16:17[music] they're they got natural marino
- 16:19wool in them, so that helps with the
- 16:20wicking. For storage, once again, the
- 16:23theme of the build is stuff that's
- 16:25really cool, but we didn't really have
- 16:26any use for it. So, I've got this Optane
- 16:28SSD. Go ahead and huck that in there.
- 16:32It's good for that to be right on the
- 16:33exhaust of a 5090, right? I think so.
- 16:37It's got to be.
- 16:40All right.
- 16:43Sure.
- 16:45Now, let's get the power supply hooked
- 16:47up. Legit excited to check this out.
- 16:49It's been years since I've unboxed a
- 16:51power supply that I'm like, whoa,
- 16:54power supply, man. But like
- 16:583,000 watts,
- 17:00what are we even talking about here?
- 17:02Now, it is important to note that when
- 17:05operating on 120 volt power, it will not
- 17:08be able to do the full 300 watts because
- 17:10you would trip your breaker. Obviously,
- 17:12in fact, this doesn't say anything about
- 17:16accepting a 120 volt input at all.
- 17:21I thought that like other, you know,
- 17:24greater than 1600 watt power supplies,
- 17:26it might at least accept both and then
- 17:28just run at a lower capacity. [sighs]
- 17:33Why did they bother to send this blah
- 17:36blah blah blah the wrench rush current
- 17:38blah blah blah DC output blah blah blah
- 17:42blah blah blah blah blah absolutely
- 17:44nothing about running at 120 volts. Oh,
- 17:48Lucas.
- 17:50Power supply question for you.
- 17:52Silverstone makes no mention whatsoever
- 17:54of 120 volt input.
- 17:58>> Yeah, this one's Yeah. 220 to 240. So, I
- 18:02wouldn't uh
- 18:03>> I mean, you could try it on 120, but it
- 18:05might uh might not go well. It's
- 18:08probably not built for it. I mean, it
- 18:10depends on the architecture. Maybe it's
- 18:12it'll work, but uh I wouldn't recommend
- 18:15it.
- 18:15>> Hi ho. Hi ho. It's back in the box it
- 18:18goes. So, we're going to go with this
- 18:20Zeus 1650R titanium. It's got dual 12vt
- 18:232x6 connectors. It should
- 18:28still maybe have enough capacity for
- 18:31this build. But just in case it doesn't,
- 18:35I am not going to install any radiators
- 18:36or anything in the bottom because hey,
- 18:39remember I can always put in a second
- 18:41power supply if it comes to that. Now we
- 18:43just
- 18:45got to get it all wired up.
- 18:47[clears throat] Okay, there's a 24 pin
- 18:50and the rest. This board takes an
- 18:53absolutely obscene number of power
- 18:55inputs. So, in addition to the 24 pin,
- 18:58we've got up to another two eight pins.
- 19:01So, quad 8 pin. I would imagine I would
- 19:03only need that if I'm overclocking the
- 19:05stuffing out of it, though. And then
- 19:07down here, we've got an 8 pin PCIe and a
- 19:10six pin PCIe. And these ones are really
- 19:12important because every one of these
- 19:15cards is allowed to draw up to 75 watts
- 19:17through the PCIe slot. Well, with seven
- 19:21slots, that could be up to 1
- 19:27>> 525 watts. That would all be over the
- 19:32like I think two 12vt wires that are in
- 19:35the 24 pin connector. So, we need these
- 19:37supplementary power connectors coming
- 19:39into the board if we're going to have a
- 19:40bunch of GPUs in here and fill this
- 19:42thing up in the longer term. Okay, here
- 19:45comes one of the 12vt 2x6s. This is not
- 19:49a super long cable.
- 19:51>> We'll make it work.
- 19:53>> There you go. Well, whatever you do,
- 19:54don't put strain on it, though.
- 19:55>> Right. Right. Right.
- 19:57>> Speaking of, uh, this is where things
- 19:59get a little tricky. As you can imagine,
- 20:00my power supply does not have a this.
- 20:04That's where this little cable I made
- 20:07for myself comes in. Nice. Right now,
- 20:11one of the cool things about 12vt
- 20:13[music] 2x6 is that it allows the power
- 20:15supply and the GPU to talk to each other
- 20:17to [music] determine how much wattage is
- 20:19needed by that GPU and whether that
- 20:21power supply can support it. So, that's
- 20:22why you'll have little markings like
- 20:24this 600 W on your cables. Um, obviously
- 20:28this contains no such communication
- 20:31pins. So, if you look closely, I've
- 20:34actually jumpered the sense wires in
- 20:37there. So, they tell it, "Yeah, pull 600
- 20:39watts. Everything's good. Don't worry
- 20:41about it."
- 20:44>> I guess you don't care about this back
- 20:46fan.
- 20:47>> What? How did that get out there?
- 20:49>> It's in between the IO.
- 20:51>> Are you kidding me?
- 20:56that. [laughter]
- 20:58[panting]
- 21:01Wait,
- 21:02>> who mounted the motherboard again?
- 21:04>> Oh, [laughter]
- 21:06it was me.
- 21:10[laughter]
- 21:11>> No,
- 21:15[laughter]
- 21:19sad.
- 21:22What's the Oh, what's the fastest way
- 21:23for us to fix this? Uh oh, man. Okay.
- 21:29>> Do we Holy balls. Okay, here's what
- 21:33we're going to do.
- 21:34>> Scissors.
- 21:39>> Do you need it?
- 21:42>> Scissors.
- 21:43>> It Oh, I hate that. Okay, I'm going to
- 21:46go get another fan and I'm going to go
- 21:48get a splitter. Find a pair of scissors.
- 21:50Don't let the people see it.
- 21:53Packing all this hardware in was [music]
- 21:55practically effortless. And it it
- 21:58doesn't even it doesn't even look that
- 22:00full. You know, it got two 5090s in
- 22:04here. No problem.
- 22:04>> I think you need two more.
- 22:06>> Well, see, you see, I did leave space
- 22:09just in case.
- 22:11I guess it's time for us to uh grab a
- 22:14monitor and look at software with some
- 22:15of the guys from the labs. Most of what
- 22:17I've been using this for lately has been
- 22:19like home assistant stuff. Obviously,
- 22:20none of this is loading because we're
- 22:22not on my local network. Um, but where
- 22:24we left off last time on software was we
- 22:26got LM Studio set up. So, we've got a
- 22:28local LLM that's running on both of the
- 22:305090s. I couldn't believe how easy it
- 22:31was to get running when I stopped
- 22:33faffing about with Adysius. Um, but you
- 22:36guys were saying that what I really need
- 22:37to be doing is running Hermes. So, I
- 22:39figured I'd bring you guys in so you can
- 22:41kind of give us a run through of what
- 22:43would be your basic starter Hermes
- 22:47setup,
- 22:48>> right? Uh, well, step one, let's install
- 22:50Hermes.
- 22:51>> Yeah, sure.
- 22:52>> You guys haven't met Stephen from the
- 22:53lab, but where would they have seen some
- 22:56of your work? Like, you're credited on
- 22:58some of the website stuff, right?
- 22:59>> Yeah, I do testing for a lot of the labs
- 23:01articles. Uh, so you see my name there.
- 23:03Uh,
- 23:04>> there you go.
- 23:04>> Frequently.
- 23:05>> He's in a Vessie ad, I think. So
- 23:07>> I am in a vessad. So I'm on the uh
- 23:10Hermes web page and they have a
- 23:12convenient install command. So I will
- 23:14run said command even though I'm not
- 23:18entirely sure what it does.
- 23:22>> Are you sure? Before we go any further,
- 23:25Lucas, do you want to give us a quick
- 23:26explanation of what Hermes is and why we
- 23:28should want it?
- 23:29>> It's just a an an agent, a harness, uh
- 23:31something like that. It's something that
- 23:32wraps around your LLM and runs it and
- 23:34kind of gives it a system prompt. makes
- 23:36other tools available and makes it like
- 23:38kind of nice to work with. Um, and then
- 23:40you know there's other integrations and
- 23:41a whole kind of ecosystem you can build
- 23:43around your LM and ties that up nicely
- 23:46hopefully in theory, you know.
- 23:48>> Yeah, I mean last time we were all
- 23:49together I think I managed to get Steven
- 23:52with three. I've never seen this before.
- 23:55My ability to break Linux is truly
- 23:57unprecedented. The whole idea for me of
- 24:00running local AI is that I am minimizing
- 24:04the amount of
- 24:06paying that I'm doing after the fact. I
- 24:08already paid a lot.
- 24:09>> Got it.
- 24:10>> So now we want to turn off the money
- 24:12spigot.
- 24:13>> Zero recurring payments. Yeah.
- 24:15>> Yeah. That
- 24:16>> for those unfamiliar, this is LM Studio.
- 24:18This is where you actually run the LM
- 24:19and it's where it's going to be hosted
- 24:21um so that it's accessible for Hermes.
- 24:23Um and but so you can host it on like
- 24:25another machine if you want on Elm
- 24:26Studio and just make it available over
- 24:27the network. Um but for here we'll just
- 24:29load the model. Uh so if we go here
- 24:32it'll load the model and it gives you a
- 24:34list of them. So Quinn 3.8 and you can
- 24:36choose your parameters and all that
- 24:37stuff.
- 24:38>> We found the default parameters were
- 24:40pretty good. I don't think there was
- 24:41pretty much anything we needed to change
- 24:42and it immediately was able to get both
- 24:44of our GPUs working together which is
- 24:46something that I had really struggled
- 24:47with with Odysius.
- 24:49>> Yeah, a lot of the defaults are just
- 24:50makes sense. And then for context, I
- 24:52think we're it's like 108k or something
- 24:54like that.
- 24:55>> Yeah.
- 24:55>> And it's estimating that's going to take
- 24:57up 25 or 29 gigs of um VRAM. So that's
- 25:01good. That'll fit on one card even
- 25:02though this is automatically uh
- 25:04splitting across two.
- 25:05>> Do we want to crank it because we have
- 25:06two cards?
- 25:07>> Uh we could crank it. Yeah. see what it
- 25:09goes up to.
- 25:10>> But that's not necessarily what might be
- 25:12right for you because if you want to run
- 25:14multiple models at the same time, each
- 25:16is going to need its own VRAMm for both
- 25:18to load the model and for its own
- 25:20context.
- 25:20>> So Hermes setup is asking us if we want
- 25:24to set up messaging. We do want to use
- 25:25Discord later, right?
- 25:26>> I do. We might as well set it up.
- 25:28>> Okay,
- 25:28>> that's one of the really cool things
- 25:30about Hermes is the interface can be as
- 25:33simple as just the chat app that you
- 25:35already have on your phone. It supports
- 25:37Discord. I believe it supports Telegram.
- 25:39>> I had a whole list. There's like Yeah.
- 25:41Telegram, Discord, a whole bunch of
- 25:42things.
- 25:42>> Yeah. So, you just have a little like
- 25:45chat. Well, yeah. For your chat bot.
- 25:49>> Yeah. I think you can like put it into
- 25:51servers or just chat to it like directly
- 25:53like direct message.
- 25:54>> Create a new application. There it is.
- 25:56Look at that. What is it going to be
- 25:58called? It's going to be called I feel
- 26:00like everyone names them Jarvis.
- 26:03>> Hermes botana.
- 26:06>> No. Sent Lana. Nice.
- 26:11>> You're assistant now. [laughter]
- 26:13>> Creative.
- 26:14>> Yeah. Yeah. I'm human for now. Maybe
- 26:18what we'll do is cut out some of the
- 26:19configuration and do a companion labs
- 26:21article. ltlabs.com. We'll have that
- 26:22link down below.
- 26:24>> Okay. I'll let you do the honors. You
- 26:26can run Hermes.
- 26:28>> Here we go. Here we go. Huge moment.
- 26:33First chat. We can fire up LM Studio and
- 26:36see if it's actually doing anything. I
- 26:38don't
- 26:38>> think it's doing anything yet.
- 26:39>> Hold on. Where's hardware?
- 26:42>> It's not doing anything.
- 26:43>> It's not doing anything. All right,
- 26:45>> let's just ask the agent.
- 26:49>> Question mark.
- 26:52>> Perfect.
- 26:56>> Yeah, I think so. I think those
- 26:57processing.
- 26:58>> It's doing stuff.
- 26:58>> It's doing things.
- 27:04This feels like in the early days of the
- 27:05internet when you'd go to like load a
- 27:08video and then you'd go make a sandwich
- 27:10and then come back
- 27:11>> the image is loading line by line.
- 27:13[laughter]
- 27:14>> This is a great example of how even
- 27:17running on something like a 5090. It's
- 27:20not really that fast. Like it's still
- 27:22just thinking. I have no output
- 27:24whatsoever.
- 27:26But if whatever it is that you're doing
- 27:29is not really timesensitive, the fact
- 27:32that it can just be sitting there
- 27:34churning away at a task can be
- 27:36incredibly [snorts] powerful. So it'll
- 27:39sit and it'll implement something and
- 27:41then test it, then troubleshoot it, then
- 27:43test it again, then troubleshoot it,
- 27:44then test it again. And all that can
- 27:46happen while I sleep. [music]
- 27:50Who needs water or electricity when we
- 27:52have this?
- 27:55After a couple small tweaks to the ENV
- 27:57file, theoretically Sventana is up and
- 28:00running. So, it's Friday. I'm kind of
- 28:03thinking I'm hosting the WAN show today.
- 28:05Any breaking tech news stories that we
- 28:07should talk about?
- 28:09Searching the web. Searching the web.
- 28:12And this is a great example of how this
- 28:14would be the sort of thing that you
- 28:16would want to schedule to run so that
- 28:17your digest is just kind of waiting for
- 28:19you in the morning.
- 28:21I realize I've said this a few times,
- 28:23but local AI is best served ahead
- 28:26[clears throat] of time.
- 28:29Can be useful, but it's not fast. Okay.
- 28:33AI agents going rogue on government
- 28:35targets. Wow, it's really heavily
- 28:38indexing AI stuff. Good lord, it's all
- 28:42AI. I feel like we have just scratched
- 28:45the very tip of the iceberg here.
- 28:48I think it does get more useful uh if
- 28:50you can get into your once you get into
- 28:51your own homes and you can integrate
- 28:53with home assistant and get it more into
- 28:55your daily routine or something like
- 28:56that and actually do instead of you know
- 28:58on our weird network and not connected
- 29:00to anything.
- 29:01>> All right. Well, I'm looking forward to
- 29:03playing around with it more and more and
- 29:05especially integration with Home
- 29:07Assistant. I want to be able to do
- 29:09things like recognize certain objects as
- 29:12they approach the house and send
- 29:14notifications or warnings. Motion
- 29:16sensors are not great because when it's
- 29:18heavily raining, I noticed last night my
- 29:19lights were all on.
- 29:20>> It could set up a rule for detect the
- 29:22weather. If it's raining, don't actually
- 29:24turn it on.
- 29:24>> Yeah. Or something similar. Um that'll
- 29:26be lower latency, but it like detect
- 29:28those patterns with the data it can take
- 29:30in. Uh
- 29:31>> super cool.
- 29:31>> Set that up for you.
- 29:33>> You know what else is cool? Telling you
- 29:35about our sponsor,
- 29:37>> Squarespace. It's the all-in-one website
- 29:39platform that helps with everything from
- 29:41building a personalized site for your
- 29:43business all the way to opening an
- 29:44online store. Create something unique to
- 29:47brand with their design intelligence
- 29:48tool. All you need to do is answer a
- 29:51couple quick questions, pick a general
- 29:53vibe, and you have a custom template. If
- 29:55you want to sell from your website,
- 29:57Squarespace can help with payment
- 29:59processing. It works with major payment
- 30:01methods like Apple Pay, Credit, Debit,
- 30:04and payment plans. They'll even assist
- 30:06with SEO tools, making sure that your
- 30:08website finds the right audience when
- 30:10people search for you online. There are
- 30:12millions of URLs to choose from to help
- 30:14with this, too. Track your performance,
- 30:17adjust on the fly, and keep growing your
- 30:19business. We've even used Squarespace
- 30:21here at LMG. Start building your
- 30:24websites today and receive 10% off your
- 30:26first purchase by visiting
- 30:28squarespace.com/ltt.
- 30:31If you guys enjoyed this video, why not
- 30:33check out the one where uh Lucas and I
- 30:35were playing around with those GPUs that
- 30:37can have a whole bunch of latent hidden
- 30:40VRAM unlocked. We've got one of them
- 30:42just like still running over there. It's
- 30:44been working great as far as I know.
- 30:46Yeah,
- 30:46>> it started with 8 gigs of VRAM and now
- 30:48it has 64.
About this transcript
This page contains the full transcript of My F-You To Sam Altman by Linus Tech Tips, generated from the public captions YouTube serves with the video. The transcript has 5,144 words across 799 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.