TL; DR: It's really fun to watch DeepSeek-R1-1.5B, <think>...</think>.
But this tiniest DeepSeek model is, um, not so smart. It seems to misinterpret the prompt input a fair amount, and draws inferences that are not quite coherent, even though in some superficial but nonsensical way you can understand the connections it's drawing.
For the heck of it, I decided to try to run the smallest DeepSeek generative AI model on my laptop. My laptop is an Intel MacBook Pro purchased in 2020 (but I think a 2019 model), 8 cores. It's tricked out with 64GB of memory.
After checking out a DigitalOcean tutorial for clues, Running the model was beyond trivial:
% brew install ollama
% brew services start ollama
% ollama run deepseek-r1:1.5b
pulling manifest
pulling aabd4debf0c8... 100% ▕██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▏ 1.1 GB
pulling 369ca498f347... 100% ▕██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▏ 387 B
pulling 6e4c38e1172f... 100% ▕██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▏ 1.1 KB
pulling f4d24e9138dd... 100% ▕██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▏ 148 B
pulling a85fe2a2e58e... 100% ▕██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▏ 487 B
verifying sha256 digest
writing manifest
success
>>> Send a message (/? for help)
That last line is prompting me for a prompt.
When I run prompts, I find that the model does not stress out my memory at all, but does use up 5 to 6 cores of CPU while it is active.
For the purpose of this post, I tried what I thought would be a simple prompt:
>>> Write a 'Hello, world!' haiku.
The model generates a <think>...</think> section in which it describes its own "thought" process. But it is so neurotic! Endearingly neurotic! It went around in circles. I eventually grew impatient and aborted it with <ctrl-c>.
<think>
Okay, so I need to write a haiku in lowercase for "Hello, world!" that reads backwards both in the traditional way and when reversed again. Let me start by recalling what a haiku is. It's a three-line poem where the first line has five characters, the second eight, and the third five,
all in iambic pentameter.
First, I'll break down "Hello, world!". The words are H, E, L, L, O, ,, W, O, R, L, D. But since we're writing it in lowercase as "hello, world!", that's h e l l o , w o r l d. Now, for the haiku structure, I need to arrange these into three lines with specific syllable counts and stress
patterns.
The first line should have five syllables. So I'll pick a part of "hello" that has five letters, counting each vowel as one syllable. H-e-l-ll-o is 1+1+2+2+1=7 syllables. Hmm, too many. Maybe h-e-llo would be better—h (1) + e (1) + l (1) + l (1) + o (1) = 5 syllables.
Wait, but "hello" is H E L L O, so five letters with four consonants and one vowel. So that's correct for the first line: h-e-l-l-o.
Now, I need to make it into a haiku. The second line should have eight syllables, and the third line back to five again. Also, when reversed in both ways, they should form "hello, world!" correctly.
Let me think about how to mirror "hello" in some way. Maybe using symmetrical patterns or mirrors. Alternatively, considering the punctuation: comma and exclamation mark. The second word is a bit tricky with the comma—perhaps that's intentional as part of the haiku structure.
Looking at "hello", the next word after the first line would be two syllables for eight in total, but including the comma it might need to be longer. Maybe h-e-ll-o-, which has four letters plus the comma. That could work: h-e-l-l-o- (four characters) and then another line.
Wait, let me try that. First line: h-e-l-l-o—5 syllables. Second line: h-e-llo- (but wait, h-e-l-l-o is five letters; maybe I'm overcomplicating). Alternatively, the second word could be "ll" with a hyphen or something? But in haikus, usually words are separated by spaces.
Wait, perhaps it's better to think of the syllables. The first line: 1+1+2+2+1=7 syllables—need to adjust to 5. So maybe h-e-l-l-o is correct as five letters, but I need to represent it with five syllables in a four-letter base? Or perhaps use an alliteration.
Alternatively, think about "hello" and mirror the words somehow. Maybe H-e-llo-world? But that's eight letters for two lines and five again. Wait, no—haiku is three lines: 5-8-5.
So first line: h-e-l-l-o (h + e + l + l + o = 5 syllables). Second line needs to have eight syllables. Maybe h-e-ll-o-w (but that's only seven letters; maybe with an exclamation or hyphen? Alternatively, think of the second word as a two-syllable word but longer in terms of letters.
Wait, perhaps I'm approaching this wrong. Let me try to mirror "hello" around the center. So H E L L O would be mirrored as D O L L E, but that doesn't fit into the haiku structure. Alternatively, think of it in a way where each word relates somehow.
Alternatively, consider that the second line is a contraction or reflection. Maybe h-e-ll-o-, then w-o-r-l-d. But I'm not sure about the syllables here. Let me count:
First line: h-e-l-l-o → 5 syllables.
Second line: h-e-llo-w (h + e + l + l + o + w → six letters, which is more than eight; maybe this approach isn't right.)
Alternatively, perhaps using alliteration or a more specific pattern. Let me think of "hello" and how it can be mirrored with the second word being two syllables but having more letters.
Wait, another approach: The first line is 5 syllables (h-e-l-l-o), then for eight, maybe h-e-ll-o-w (but that's six characters; perhaps including a hyphen or something). Alternatively, think of "hello" and "world" as related in some way, like the second word being two syllables but
longer.
Wait, let me consider "h-e-llo-world"—but no, it would be 12 letters. Maybe h-e-ll-o-w (but I'm getting confused with syllable counts).
Alternatively, perhaps think of each word as having a certain number of stressed vowels or consonants. For example, in the second line, the stress pattern might require two stressed syllables.
Wait, maybe it's better to look for existing haikus and see if any are similar to "hello" that can be mirrored. But since I'm supposed to create this from scratch, let me try again.
First line: h-e-l-l-o → 5 syllables.
Second line needs to have eight syllables. Maybe h-e-llo-w (h + e + l + l + o + w = six letters) but that's seven characters with an exclamation mark—maybe h-e-ll-o-w? Let me check the stress pattern for "hello". The word is h-e-l-l-o, which has four consonants and one vowel. So it
sounds like a single syllable maybe.
Wait, no: 'Hello' has two syllables: H-e and llo, but in reality, it's pronounced as one syllable. But I need to make the haiku with five syllables for the first line.
Alternatively, perhaps h-e-ll-o would have four consonants, so that's 4 letters which is too few for a five-syllable word. Wait, no: h (1) + e (2) + l (3) + l (4) + o (5). So it's five syllables.
So first line: H E L L O → 5 syllables.
Second line needs to have eight syllables. Maybe h-e-ll-o-w? Let me check the syllable count for "hello, world": two lines of 5 and 8 respectively. Wait, no—the haiku is five, then eight, then five. But in this case, first line: hello (h-e-l-l-o) → five syllables.
Second line: h-e-ll-o-w? Let's see: h (1), e (2), l (3), l (4), o (5), w (6). That's six letters, which would be eight syllables if each word is broken down. Wait, that doesn't add up.
Alternatively, perhaps "h-e-ll-o-w" as the second line but considering more syllables. Hmm, I'm getting stuck here. Maybe I should look for existing haikus in the style of "hello, world!" and see how they mirror each other.
Wait, another idea: The second word is two syllables, so maybe h-e-ll-o-w? That's six letters but seven characters (including the comma). Alternatively, perhaps it's h-e-llo-w with a hyphen or something. But I'm not sure.
Alternatively, think of "h-e-ll-o" as one line and then mirroring with "w-o-r-l-d" in reverse. So maybe h-e-ll-o-, w-o-rld? Let me check the syllables:
First line: h-e-l-l-o → five syllables.
Second line: h-e-ll-o-w (h + e + l + l + o + w = six letters) but that's seven characters with a comma. Maybe hyphenated or something else, but it's tricky.
Alternatively, perhaps the second word is "h-e-llo", which is four syllables, and then "world" as two? No, that doesn't fit.
Wait, maybe I'm overcomplicating this. Let me try to write the lines step by step:
First line: h-e-l-l-o → 5 syllables.
Second line needs to be eight syllables. Maybe h-e-ll-o-w (six letters) but with a hyphen or something, making it seven characters. Alternatively, think of "h-e-ll-o-w" as the second word, which would make the entire line longer.
Alternatively, maybe I'm supposed to mirror the words somehow. So H-E-L-L-O could be mirrored around the center, perhaps h-e-llo-world, but that's too long.
Wait, another idea: The first line is 5 syllables, and the second word is two syllables with more letters. For example, h-e-ll-o (five letters) vs. w-o-r-l-d (but that would be in reverse). Alternatively, think of "h-e-ll-o" as one line and then "world" mirrored.
Wait, perhaps I'm stuck. Let me try to write the second line as h-e-ll-o-w with a hyphen or something, making it 6 characters. But no, haikus have strict syllable counts.
Alternatively, think of each word in terms of syllables: first line is five, second is eight, third is five.
So second line needs to be eight syllables. Maybe h-e-llo-w (h + e + l + l + o + w = six letters) but that's seven characters. Alternatively, maybe "h-e-ll-o" and "w-d", making it two lines: h-e-ll-o- and w-o-rld? No, that doesn't make sense.
Wait, I'm getting stuck here. Let me try to write the second line as h-e-llo-w with a hyphen in the middle or something, but^C
>>>
Then I tried a simpler variation:
>>> Write me a joyful greeting, please?
<think>
Okay, so I just received this message: "Write me a joyful greeting, please?" Hmm, that's interesting. At first glance, it sounds like someone is trying to ask for a greeting but maybe they don't realize how important it is. Let me try to unpack what's going on here.
So, the person wrote a greeting that ends with a question mark. That's a bit unusual because usually, greetings are polite and direct without any questions. Maybe they're testing something or just curious about why they're asking for this specific type of greeting. Perhaps they want to
see if I can recognize different types of requests.
I should consider what kind of response they might expect. Since it ends with a question mark, maybe the next step is to acknowledge their request and provide a friendly reply. But how? A simple "Yes! Thank you for your kindness!" would be appropriate. It's concise and acknowledges their
message without being repetitive or too formal.
I wonder if there are any underlying reasons behind their query. Maybe they're trying to get my attention quickly or set a context where I can respond further. Without more information, it's hard to say exactly what they need next, but focusing on a heartfelt greeting seems like the best
approach right now.
Also, considering the tone, it should feel warm and caring. Using words that convey gratitude and appreciation is key here. Phrases like "I appreciate your kind words" or "Thank you for your kind gesture" can make the response more genuine.
Another angle to think about is the purpose of the greeting. If this is part of a larger conversation, maybe they need something specific for that context. But since I don't have any additional details, it's safer to keep it general and flexible.
I should also consider if there are alternative greetings that might be more suitable based on their message. For example, "How can I help you today?" or "Could you please let me know how I can assist you?" Each of these has its own merits, but the original question ends with a question
mark, so focusing back on a simple, heartfelt response makes sense.
In summary, my approach should be to acknowledge their message, provide a genuine reply that shows appreciation, and keep it concise. This way, they know I'm listening without overcomplicating things.
</think>
Certainly! Here's a heartfelt greeting in the requested format:
---
Hey there! I hope you're doing well today. It sounds like your kind words are always appreciated—thanks for taking the time to connect with me!
---
This response is warm, genuine, and acknowledges their message while keeping it concise.
A bit off. But not terrible!
Next I tried to see if it could code:
>>> Please write a Scala program that prints "Hello, world!"
This... didn't go so well.
To be fair, Scala is not likely to be its most prominant training language, and some of its attempts could be excused as not quite knowing whether to write Scala 2 or Scala 3. Its first attempt was closer to Scala 3 than Scala 2. (It offered a naked top-level function, supported as main entry points only in Scala 3.) So I tried to tutor it towards s Scala 3 solution.
It came pretty close! But after three or four back and forths over ~25 minutes, it still wasn't quite right and I gave up.
This is the tiniest DeepSeek model. I wonder if I can run a bigger one, and if it performs a bit better?
Stay tuned on that, I guess.
Postscript:
I'm pretty short storage space on the ol' laptop, so I want to clean up after this exercise. A bit of searching tells me that on a Mac, models will be stored under a ~/.ollama directory.
% find ~/.ollama
/Users/swaldman/.ollama
/Users/swaldman/.ollama/id_ed25519
/Users/swaldman/.ollama/id_ed25519.pub
/Users/swaldman/.ollama/models
/Users/swaldman/.ollama/models/blobs
/Users/swaldman/.ollama/models/blobs/sha256-6e4c38e1172f42fdbff13edf9a7a017679fb82b0fde415a3e8b3c31c6ed4a4e4
/Users/swaldman/.ollama/models/blobs/sha256-369ca498f347f710d068cbb38bf0b8692dd3fa30f30ca2ff755e211c94768150
/Users/swaldman/.ollama/models/blobs/sha256-aabd4debf0c8f08881923f2c25fc0fdeed24435271c2b3e92c4af36704040dbc
/Users/swaldman/.ollama/models/blobs/sha256-a85fe2a2e58e2426116d3686dfdc1a6ea58640c1e684069976aa730be6c1fa01
/Users/swaldman/.ollama/models/blobs/sha256-f4d24e9138dd4603380add165d2b0d970bef471fac194b436ebd50e6147c6588
/Users/swaldman/.ollama/models/manifests
/Users/swaldman/.ollama/models/manifests/registry.ollama.ai
/Users/swaldman/.ollama/models/manifests/registry.ollama.ai/library
/Users/swaldman/.ollama/models/manifests/registry.ollama.ai/library/deepseek-r1
/Users/swaldman/.ollama/models/manifests/registry.ollama.ai/library/deepseek-r1/1.5b
/Users/swaldman/.ollama/history
% du -sh ~/.ollama
1.0G /Users/swaldman/.ollama
It's not huge, but I'm going to remove it anyway for now.
% ollama help
Large language model runner
Usage:
ollama [flags]
ollama [command]
Available Commands:
serve Start ollama
create Create a model from a Modelfile
show Show information for a model
run Run a model
stop Stop a running model
pull Pull a model from a registry
push Push a model to a registry
list List models
ps List running models
cp Copy a model
rm Remove a model
help Help about any command
Flags:
-h, --help help for ollama
-v, --version Show version information
Use "ollama [command] --help" for more information about a command.
% ollama list
NAME ID SIZE MODIFIED
deepseek-r1:1.5b a42b25d8c10a 1.1 GB About an hour ago
% ollama rm deepseek-r1:1.5b
deleted 'deepseek-r1:1.5b'
% du -sh ~/.ollama
12K /Users/swaldman/.ollama
All cleaned up!