Tales from the jar side: DeepSeek R1, Semester begins, jChampions Conf, and the usual silly toots and skeets

I started a band called Duvet, which only plays covers. (rimshot, c/o @dak.bsky.social)

Welcome, fellow jarheads, to Tales from the jar side, the Kousen IT newsletter, for the week of January 19 - 26, 2025. This week I taught my Deep Dive Into Spring course as an NFJS virtual workshop, and began my new semester at Trinity College (Hartford).

DeepSeek R1

The big news in the AI world this week was the release of DeepSeek R1, an open source model from a company called DeepSeek-AI. That’s a big deal, for several reasons:

• The company is in China. Given the current controversy over tools like TikTok and other potentially Chinese government-controlled products, that’s significant. After all, only our own companies are allowed to spy on us and steal our data, not the Chinese!

• The US spent the last few years deliberately trying to make it difficult for China to acquire the high-powered GPU chips needed for giant AI models. Apparently, in this case at least, that didn’t matter.

• Unlike OpenAI and its commercial competitors, the DeepSeek model is completely free and open source, under the MIT license, which means you can pretty much do whatever you want with it. It’s got the normal guard rails you might expect from China, but since it’s open source, somebody will create a fork without those limitations soon enough.

• It’s a reasoning model competing with OpenAI’s o1 (and upcoming o3) models. Not only does it beat them, it supposedly was created for a fraction of their cost, and it’s free to use.

In the AI world, there’s free and there’s free. Sure, you can download the model weights and run it locally (I’ll get back to that), and apparently you can run their model through the web interface for free as well, though I had issues accessing it.

I went to the API, however, and saw this:

Pricing Notice
1. The deepseek-chat model will be charged at the discounted historical rate until 16:00 on February 8, 2025 (UTC). After that, it will be charged at $0.27 per million input tokens and $1.10 per million output tokens.
2. The deepseek-reasoner model will launch with pricing set at $0.55 per million input tokens and $2.19 per million output tokens.

Those prices are actually comparable to OpenAI, or significantly lower, depending on what you do. I may have done the math incorrectly, but DeepSeek looks like a really good deal.

To access the API, you need to sign up for an API key. That’s only allowed via email, and when I put in my normal kousenit.com email address, I got an error saying that it wasn’t eligible for some reason. I therefore went with my Trinity College email address instead, and they were fine with that. I don’t why exactly, but at least I’m in.

The next step is to use their API, and, lo and behold, they too are piggy-backing on OpenAI’s Python API. That’s getting to be a standard for the industry, especially for AI companies that don’t want the trouble of creating their own API. (I made a whole video about Google’s Gemini being the latest to support it.) All you need to do is use their endpoint and their API key, and you’re ready to run.

Here’s my first call, based on their demo:

Want to see how a reasoning model reasons? Here’s the response I got:

Got there eventually, I suppose, but it took a while.

That’s not the special reasoning model (which they call deepseek-reasoner), but it’s still using chain-of-thought reasoning, which basically means show your work as you go. the result is pretty detailed, but at least it got the right answer.

I also tried the free model, which is available via Ollama.

See those numbers at the bottom? Those are the number of parameters, in billions. On my spiffy new Mac M4 hardware, I can run the 14b and 32b parameter models with ease, and the 70b parameter without much trouble, but I wouldn’t even attempt the massive 671b model. That would bring my poor laptop to its knees.

With the 70b model, I asked the “r’s in strawberry” question and got quite the fiasco. I won’t dump it here, but it went on and on, knowing something was wrong but being unable to do much about it, until it concluded there were only two r’s in the word strawberry. It was quite a spectacle to see, to be honest.

I’m not sure what I’m going to do with all this, other than show my students in my AI class. The big threat is to OpenAI. Sam Altman is going have trouble funding his little $500 billion Stargate project if the Chinese can compete with it at a tiny fraction of the cost and then give away the result. You’re probably aware that OpenAI has never come close to being profitable. If anything, this shows just how doomed they are.

Hey, I like Claude better anyway. :)

New Semester Begins

Speaking of Trinity College, this semester I’ve got a couple of courses:

• Artificial Intelligence

• Senior seminar, part 2

That’s along with managing several senior projects, as well as a couple of independent study ones.

I have to make a comment about the AI course. When you look at the description in the course catalog, it’s seriously dated. Of course, anything older than about three years in this industry would be. I’m using the Manning book Build a Large Language Model (From Scratch) in the course, which I believe I mentioned in an earlier newsletter.

On the No Fluff, Just Stuff tour, I often wrote abstracts for what I called aspirational talks, meaning I intended to give them but was in no way qualified to do so at the time I proposed them. My current AI course probably fits into that category. I mean, I know something about the subject, but I’m going to be figuring out a lot as I go along. I’m also a mediocre Python coder at best, so I’ll be learning that throughout the semester as well. Like so many similar items, this is good for me, but will be stressful along the way.

(If you’re one of my students and for some reason you’re reading this, you might want to consider dropping the course and taking it next time around, when I’ve got it figured out a lot better. Or be prepared for some level of chaos. Up to you either way.)

jChampions Conference

The annual conference presented by members of the Java Champions community is called jChampions Conference, and ran online, for free, last Thursday and Friday and continues this upcoming Monday and Tuesday.

I gave a talk on Friday afternoon called Data-Oriented Programming and REST APIs:

That’s the link to the live stream on their channel. Since the group used Streamyard as their platform and I have an account there, I was able to stream it to my own Tales from the jar side YouTube channel simultaneously.

The most amazing thing about my presentation, and I can’t believe I’m saying this, is that I actually finished early. I ran through all the examples I had and answered all the questions, and still had about 5 minutes to go. Imagine that. It can be done.

Here’s the GitHub repository with the samples if you want it.

I have several conflicts Monday and Tuesday, but I hope to watch a few sessions anyway if I can. Of course, as you can see, they’re all available on the YouTube channel as they’re given.

Toots and Skeets

Good question

I’m assuming they’re just called tribbles, but the related question is, what do they evolve into? That’s a scary thought.

Lego instructions

Good luck with that.

I wonder why it does that?

I hate when that happens. On a related note:

Hard to argue with it, though.

Please listen because…

Now I’m going to wonder about that every time I reach an automated phone system.

And I helped

That falls into the “can’t believe that never occurred to me” category of jokes.

It’s a trap!

I feel your pain, buddy, but start with, “you look more beautiful today than the day I married you” and go from there.

Related to this week’s subtitle

Come on baby light that joke on fire.

I’m not the only one with really old song references

I suppose the question to ask is, were you wearing a red hat at the time? Apparently that matters.

Old movie references, too

If the zoo doesn’t play Also Sprach Zarathustra over the cage speakers they’re missing an opportunity.

Probably should have been this week’s subtitle

Yes. Yes it does. And finally:

Another one bites the dust

Have a great week, everybody!

Last week:

• Deep Dive Into Spring, an NFJS Virtual Workshop

• New semester at Trinity College begins

This week:

• Managing Your Manager, on the O’Reilly Learning Platform

• LangChain4j, an NFJS Virtual Workshop

• Functional Java, on the O’Reilly Learning Platform

• I’m scheduled for jury duty on Friday. Hopefully it will cancel, but I’ll let you know.