DeepThought-8B: A small, capable reasoning model

There&#x27;s been a rush of releases of reasoning models in the past couple of weeks. This one looks interesting, too.<p>I found the following video from Sam Witteveen to be a useful introduction to a few of those models:<p><a href="https:&#x2F;&#x2F;youtu.be&#x2F;vN8jBxEKkVo" rel="nofollow">https:&#x2F;&#x2F;youtu.be&#x2F;vN8jBxEKkVo</a>

In what way did they &quot;release&quot; this? I can&#x27;t find it in hugging face or ollama, and they only seem to have a &quot;try online&quot; link in the article. &quot;Self-sovereign intelligence&quot;, indeed.

Legally, you cannot name the llama3 based models like that, YOu have to use, llama in the name<p><a href="https:&#x2F;&#x2F;huggingface.co&#x2F;meta-llama&#x2F;Llama-3.1-70B-Instruct&#x2F;blob&#x2F;main&#x2F;LICENSE" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;meta-llama&#x2F;Llama-3.1-70B-Instruct&#x2F;blo...</a>

Am I wrong to think that &quot;reasoning model&quot; is a misleading marketing term?<p>Isn&#x27;t it a LLM with an algo wrapper?

Given the name they gave it, someone with access should ask it for the “Answer to the Ultimate Question of Life, The Universe, and Everything”<p>If the answer is anything other than a simple “42”, I will be thoroughly disappointed. (The answer has to be just “42”, not a bunch of text about the Hitchhikers Guide to the Galaxy and all that.)

&quot;what is the population of manhattan below central park&quot;<p>ChatGPT-o1-preview: 647,000 (based on 2023 data, breaking it down by community board area): <a href="https:&#x2F;&#x2F;chatgpt.com&#x2F;share&#x2F;674b3f5b-29c4-8007-b1b6-5e0a4aeaf0e9" rel="nofollow">https:&#x2F;&#x2F;chatgpt.com&#x2F;share&#x2F;674b3f5b-29c4-8007-b1b6-5e0a4aeaf0...</a> (this appears to be the most correct, judging from census data)<p>DeepThought-8B: 200,000 (based on 2020 census data) Claude: 300-350,000 Gemini: 2.7M during peak times (strange definition of population !)<p>I followed up with DeepThought-8B: &quot;what is the population of all of manhattan, and how does that square with only having 200,000 below CP&quot; and it cut off its answer, but in the reasoning box it updated its guess to 400,000 by estimating as a fraction of land area.

I asked it &quot;Describe how a device for transportation of living beings would be able to fly while looking like a sphere&quot; and it just never returned an output

The reasoning steps look reasonable and the interface is simple and beautiful, though Deepthought-8b fails to disambiguate the term &quot;the ruliad&quot; as the technical concept from Wolfram physics, from this company&#x27;s name Ruliad. Maybe that isn&#x27;t in the training data, because it misunderstood the problem when asked &quot;what is the simplest rule of the ruliad?&quot; and went on to reason about the company&#x27;s core principles. Cool release, waiting for the next update.

I am very impressed. I asked chat.ruliad.co<p><pre><code> Beginning from the values for fundamental physical constants, is it possible to derive the laws of entropy ? </code></pre> and then based on its response to that I asked it<p><pre><code> Based on this analysis, can you identify and describe where the dividing line is between (a) the region where (microscopic&#x2F;atomic) processes are reversible, and (b) the region where macroscopic processes are irreversible ?</code></pre>

Is the source code available for this? And who is behind the company?

I asked it &#x27;find two primes whose sum is 123&#x27; .. it is in deep thought from 5 minutes just looping and looping over seemingly repeated hallucinations of right path. (btw, chatgpt immediately answers 61 and 62 lol.. so much for intelligence)

It <i>looks</i> nice, but my chrome browser on android has all sorts of trouble with rendering the animated bits, so it ends up skipping frames throughout my navigating and clicks. Add to that; the model doesn&#x27;t respond at all in my multiple attempts, it&#x27;s a waste of time until it&#x27;s remedied.

Is it possible to try it without logging in?<p>Can you log in with anything other than a Google account?<p>I was excited by the tagline &quot;Self-Sovereign&quot;, but it appears this is not.

&quot;Model A 13B&quot;, &quot;Model B 20B&quot; etc are pretty vapid claims. Which <i>actual models</i>? There are plenty of terrible high-param-count models from a year or two ago. The benchmark seems meaningless without saying what models are actually being compared against... And &quot;13B&quot; in particular is pretty sketchy: are they comparing it against <i>Llama 2 13B</i>? Even an untuned Llama 3.1 8B would destroy that in any benchmark.<p>Smells a little grifty to me...

It’s just a web page. How to try the model?

A bit off-topic, but that comparison graph is a great example why you should buy your designer a cheap secondary screen. I was viewing it on my second monitor and had to lean in to make out the off-white bar for Model D on the light-grey background. Moved the window over to my main screen and it&#x27;s clear as day, five nice shades of coffee on a light-gray background.

[flagged]

Not bad, asked it to count Rs in &quot;strawberry&quot; and Ns in &quot;international&quot;, it answered correctly, and it was fast.