WEBVTT

00:00:00.900 --> 00:00:03.102
Episode 4 said: leave it on.

00:00:03.493 --> 00:00:04.604
But by default,

00:00:04.654 --> 00:00:08.618
a plugged-in Windows 11 PC
can sleep after five minutes.

00:00:09.839 --> 00:00:12.769
And updates can restart
it outside active hours:

00:00:12.980 --> 00:00:15.246
five in the evening
to eight in the morning.

00:00:17.200 --> 00:00:22.679
AI for Research Efficiency.
Episode 14: Where your agent lives.

00:00:24.300 --> 00:00:26.867
Last time,
you saw how this series is made.

00:00:27.589 --> 00:00:29.837
This time: where an agent can live,

00:00:29.887 --> 00:00:32.475
and how to keep it running
and reach it safely.

00:00:34.100 --> 00:00:35.794
One: the rungs.

00:00:36.730 --> 00:00:39.834
In a browser, the assistant
works on copies you upload.

00:00:40.165 --> 00:00:44.551
An agent on your laptop works in your
own folders, while the laptop is awake.

00:00:45.700 --> 00:00:47.545
The owner of this
series went further:

00:00:47.916 --> 00:00:52.668
first a small always-on computer, for
long runs steered from somewhere else.

00:00:53.768 --> 00:00:59.184
The owner's was a Beelink SER8, a
mini PC with 16 gigabytes of memory,

00:00:59.314 --> 00:01:01.219
known on the network as newt.

00:01:02.380 --> 00:01:05.542
Then a server with local tools,
in the owner's words,

00:01:05.732 --> 00:01:09.344
to help ensure both strict
confidentiality and cost control.

00:01:09.694 --> 00:01:12.795
Each rung adds something,
and costs something.

00:01:14.900 --> 00:01:16.665
Two: leave it on.

00:01:17.399 --> 00:01:21.001
OpenAI's own documentation
suggests a machine like this:

00:01:21.492 --> 00:01:25.884
use a dedicated always-on Mac
or Windows PC when you want ChatGPT

00:01:25.955 --> 00:01:28.436
to stay reachable
for longer-running work.

00:01:29.632 --> 00:01:31.754
The checklist, Windows first.

00:01:32.264 --> 00:01:37.238
Sleep: when plugged in, set
Make my device sleep after to Never;

00:01:37.589 --> 00:01:40.521
on a Mac,
turn on Prevent automatic sleeping.

00:01:40.812 --> 00:01:44.454
And on a laptop, closing the
lid can still put it to sleep.

00:01:45.574 --> 00:01:48.760
Some agents keep it awake
while they run: Antigravity,

00:01:48.930 --> 00:01:54.140
Claude Code on a Mac, and Codex with an
experimental setting, off by default.

00:01:55.360 --> 00:01:59.986
Updates: stretch active hours
over your runs, up to 18 hours,

00:02:00.156 --> 00:02:02.309
and restart it yourself between runs.

00:02:02.680 --> 00:02:06.114
After a power cut,
a Mac desktop can start up by itself,

00:02:06.324 --> 00:02:08.346
and so can some mini PCs.

00:02:10.100 --> 00:02:12.595
Three: keep the session alive.

00:02:13.356 --> 00:02:17.962
Close a terminal, and the agent in it
stops; each tool has a way around that.

00:02:18.233 --> 00:02:20.986
Claude Code's background
sessions need no terminal.

00:02:21.197 --> 00:02:25.002
They're a research preview,
and a shutdown or restart stops them.

00:02:25.882 --> 00:02:28.785
Antigravity's remote control
can run as a service;

00:02:29.035 --> 00:02:32.498
on Windows, starting it with the
computer needs an administrator.

00:02:33.443 --> 00:02:37.307
Codex's always-on host is
the ChatGPT desktop app.

00:02:37.638 --> 00:02:43.624
And on a Mac, Linux, or WSL, tmux keeps
an agent running after you disconnect.

00:02:44.584 --> 00:02:49.347
Cloud scheduled tasks, in Claude Code
and Codex, run with your computer off,

00:02:49.517 --> 00:02:53.980
on paid plans. Anything that needs
your own files needs the machine on.

00:02:55.700 --> 00:02:58.119
Four: reach it from anywhere.

00:02:58.900 --> 00:03:00.702
Each agent has a remote control.

00:03:01.072 --> 00:03:03.605
Claude Code's
and Codex's need a paid plan;

00:03:03.935 --> 00:03:06.758
Antigravity's documentation
names no plan.

00:03:07.678 --> 00:03:12.107
For everything else, there's SSH,
a terminal on the other machine,

00:03:12.217 --> 00:03:13.309
over Tailscale,

00:03:13.580 --> 00:03:16.966
a private network that works without
opening ports on your router.

00:03:17.887 --> 00:03:21.957
The vendors give two cautions: don't
expose it directly to the internet,

00:03:22.188 --> 00:03:24.955
and sign in with keys, not passwords.

00:03:25.895 --> 00:03:30.497
And Windows Remote Desktop needs
Windows Pro on the PC you connect to.

00:03:31.537 --> 00:03:36.766
The owner's own route changed over time:
first Remote Desktop; then WireGuard,

00:03:36.836 --> 00:03:40.192
the protocol Tailscale
is built on, and SSH;

00:03:40.623 --> 00:03:44.709
now an SSH app,
with tmux in several panes.

00:03:46.100 --> 00:03:48.118
Five: the power bill.

00:03:48.899 --> 00:03:49.760
In Arkansas,

00:03:49.911 --> 00:03:53.588
home power averaged about
13.7 cents a kilowatt-hour

00:03:53.638 --> 00:03:56.363
from January to August 2026.

00:03:57.640 --> 00:04:03.107
Left on all year, 10 watts costs
about $12, and 50 watts about $60.

00:04:03.578 --> 00:04:06.211
300 watts costs about $360:

00:04:06.622 --> 00:04:10.186
some $30 a month,
more than a $20 plan.

00:04:11.326 --> 00:04:14.068
Apple lists four watts at idle
for the Mac mini.

00:04:14.479 --> 00:04:17.261
For other machines,
a plug-in meter settles it.

00:04:19.700 --> 00:04:22.506
Six: a server with local models.

00:04:23.388 --> 00:04:25.832
The owner's server runs
its own language model,

00:04:25.962 --> 00:04:29.707
transcription and speech,
on two large graphics cards.

00:04:30.648 --> 00:04:33.269
It cost more than $30,000.

00:04:33.759 --> 00:04:38.211
By the owner's estimate, it has
saved more than 360 hours to date,

00:04:38.321 --> 00:04:39.752
over roughly nine months,

00:04:40.122 --> 00:04:43.494
and made possible projects like
video generation and testing;

00:04:43.924 --> 00:04:47.966
it's also used for editing and improving
internal grant development work.

00:04:48.946 --> 00:04:50.407
How good are local models?

00:04:50.857 --> 00:04:56.532
On one public index, in October 2026,
the best closed model scored 58;

00:04:56.842 --> 00:05:02.156
the owner's model, 40; and models
that fit one workstation, 12 to 34.

00:05:02.606 --> 00:05:06.959
Good enough, in our reading, for bulk
reading, transcription and first drafts;

00:05:07.370 --> 00:05:10.131
not a replacement
for the frontier on hard work.

00:05:11.112 --> 00:05:15.357
Open weights don't share one license:
Gemma 4's is Apache,

00:05:15.708 --> 00:05:19.282
a Qwen research license allows
research or evaluation only,

00:05:19.633 --> 00:05:22.936
and Llama 4 asks you
to display Built with Llama.

00:05:23.927 --> 00:05:26.249
And local isn't the same as approved.

00:05:26.620 --> 00:05:28.111
At the University of Arkansas,

00:05:28.301 --> 00:05:32.495
any AI tool used on
a university system

00:05:32.705 --> 00:05:34.587
or for university business
must be vetted and approved.

00:05:35.527 --> 00:05:40.895
NIH's ban on AI in review critiques
names the technology, not where it runs,

00:05:40.945 --> 00:05:42.157
in our reading.

00:05:42.207 --> 00:05:45.422
Health data and controlled
unclassified information bring rules

00:05:45.472 --> 00:05:46.773
for the machine itself.

00:05:47.853 --> 00:05:49.675
One trap: in Ollama,

00:05:49.926 --> 00:05:54.511
a model whose name ends in
:cloud runs on Ollama's servers.

00:05:55.700 --> 00:05:58.598
Seven: your agent and a local model.

00:05:59.400 --> 00:06:03.609
Codex documents it: --oss,
with Ollama or LM Studio.

00:06:03.720 --> 00:06:06.476
Claude Code works through a
setting those two document,

00:06:06.526 --> 00:06:09.993
but loses its remote control.

00:06:11.093 --> 00:06:15.602
Antigravity's command-line tool
documents only Google's Gemini API;

00:06:15.993 --> 00:06:18.898
its Python SDK can run local models.

00:06:20.124 --> 00:06:22.667
So, a pattern
that keeps the data local:

00:06:23.017 --> 00:06:26.511
a frontier agent builds and tests
the process on a made-up file

00:06:26.561 --> 00:06:31.425
with the same columns. Then you run it
on the real data, with a local model.

00:06:34.100 --> 00:06:36.681
Eight: two ways not to buy.

00:06:37.463 --> 00:06:40.307
Cloud agents keep working
while your laptop sleeps:

00:06:40.677 --> 00:06:43.451
Claude Code's
and Codex's on paid plans,

00:06:43.681 --> 00:06:47.086
and Google's Jules,
free for 15 tasks a day.

00:06:47.936 --> 00:06:49.959
But they work on a
copy of your project,

00:06:50.029 --> 00:06:53.413
in the vendor's cloud,
not on the files on your computer.

00:06:54.244 --> 00:06:55.625
Or use your campus.

00:06:56.016 --> 00:06:59.790
Most of the Arkansas High Performance
Computing Center's services are free

00:06:59.840 --> 00:07:04.325
to eligible researchers,
with GPU jobs of up to 72 hours.

00:07:05.296 --> 00:07:10.460
Nationally, NSF's ACCESS costs nothing,
and you don't need an NSF award.

00:07:10.771 --> 00:07:15.655
The National AI Research Resource pilot
requires open, publishable results.

00:07:17.300 --> 00:07:20.262
Nine: so,
where should your agent live?

00:07:21.058 --> 00:07:26.028
Runs that outlast the lid: a small
always-on computer, or a cloud agent.

00:07:26.359 --> 00:07:31.328
Graphics cards now and then: your campus
cluster or ACCESS, before you buy.

00:07:32.299 --> 00:07:35.686
Confidential data:
your campus's approved route first,

00:07:35.856 --> 00:07:38.421
and a local model only
with that approval.

00:07:39.321 --> 00:07:43.969
Wherever it lives, secure it: updates
on your schedule, an encrypted disk,

00:07:44.200 --> 00:07:47.305
nothing open to the internet,
and keys to sign in.

00:07:48.500 --> 00:07:50.001
Pause here, and try it:

00:07:50.331 --> 00:07:53.764
ask your agent to run the always-on
check from the companion page.

00:07:54.034 --> 00:07:56.065
You should see a
short report on sleep,

00:07:56.236 --> 00:07:59.398
updates and remote access,
with nothing changed.

00:08:02.700 --> 00:08:06.810
So: keep it awake, keep the
session alive, reach it safely,

00:08:07.021 --> 00:08:11.271
count the power, and remember that
local isn't the same as approved.

00:08:14.700 --> 00:08:15.562
That's the route.

00:08:15.973 --> 00:08:21.285
Next: In Practice, starting with Ask
for the process, not the answer.