Over the last week of December 2024, there was a new foundational model eponymously named Deepseek v3. The model is available under the MIT License which is excellent. The training data has not (yet?) been released, so technically, we don’t have any way to understand what it was trained on and won’t be able to reproduce the model.

Not that it is a showstopper.

Intrigued by the reviews of the release, I tried their chat offering. I was asked to set up an account on chat.deepseek.com. I used my default email ID (that of @ieee.org) but it refused to let me register, saying that domain is not supported! That’s quite a claim. Why was it a problem, I have no idea.

I used another email ID and that worked.

With that sorted out, the UI presented was similar to any of the others we’ve seen so far.

I posed a bunch of prompts and the answers were streamed back very rapidly and more or less accurate.

I loved how it replied to this prompt: “1^3+2^3+3^3+4^3+5^3+6^3+7^3+8^3+9^3 =?”. It came back in Chinese (hmm) and when asked to respond in English, it was beautiful to watch it construct the equations. It provided me the formula for computing the sum of cubes, wonderful.

By contrast, the replies from Gemini and ChatGPT was mundane and boring.

After some contemplation, I asked this: “I was in Beijing on 4 june 1989. what would i have seen happening in the capital?”.

I prompted that to Deepseek, Gemini and ChatGPT. Here’s what I got back:

Both Gemini and ChatGPT gave the historical information, Deepseek refused and wanted to move on.

How would one ascertain if the foundational models being used have information that has been censored especially if the censorship was because of political reasons?

It is one thing to say that Molotov cocktail making recipes won’t be provided no matter what, but it is another to completely skip historical events because it is politically inconvenient, or sensitive or deemed unacceptable by the political class.

I do really want to see Deepseek excel in what they are doing but these stumbles do not help in raising confidence or credibility. These actions feed into the unneeded narrative that one can’t trust technology from China.

That does not need to be the case.

As it is, we understand that all foundational models hallucinate and we manage accordingly. But for models to refuse to answer politically sensitive questions, that’s absolutely unacceptable.

Foundational models are increasingly used in many fields and without any knowledge of the training data. The resultant model that one uses could be exposing your generative AI applications to issues that you have no control over. This is not crying wolf – it is a legitimate call for full transparency for the good of everyone, everywhere.

Over the last 30-odd years in the software world, we have seen the enormous success of free and open source software – Linux, Apache, Nginx, LibreOffice, Firefox, MariaDB, Python, TensorFlow, Pytorch etc. Open source technology is used in all aspects of businesses and consumer settings. OSS solves problems at hand and does it well enough and continues to deliver. And all of this was achieved not by chasing after any “market capitalization” considerations.

Open source removes the barriers to participate for anyone, anywhere. Armed with a computer and an Internet connection, anyone can be a contributor, participate in any open source project and, of course, consume these technology.

It is fair to say that open source software adoption happened with minimal use of marketing (mostly by commercial open source companies and foundations like Red Hat, SUSE, Nextcloud, Eclipse, Linux Foundation, CHAOSS etc).

As we get into the third year of the Generative AI era, we need to recognize that when we talk about open source in the AI space, it is NOT only about the source code of the application (in this case the foundation model), but also how the model was created via the data that was used to train the model.

Denying access to the training data for both inspection and review diminishes the credibility and trustworthiness of the resulting model.

We see that clearly in the Deepseek example above. We still do not have access to the training data (yet) and perhaps the reason for refusing to answer questions around June 1989 is probably hiding there. We do know that the Chinese government vets the training data set to weed out data that they deem unacceptable.

The Open Source Iniitiative released the Open Source AI Definition in October 2024.

While OSI maintains and is the custodian of the open source licenses, the Open Source AI Definition from the OSI, from my perspective, falls short of Four Freedoms needed as per the Free Software definition that OSI has adopted and expounds.

The source, in the case of any generative AI models, is the data. The data is the starting point based on which the model is trained. Once trained, users engage with it directly via user prompts, APIs, system prompts or by finetuning doing things like RAGs (retrieval augmented generation) etc.

It is completely understandable that training data might have challenges in being released – and no one is forcing it to be released anyway.

The training data could be medical information, financial information or something that, even if made public under an open data commons license, might be unacceptable because it could contain personally identifiable information (PII). PII that could not be fully anonymized.

Under those conditions, I will assert that the resulting model is NOT open source compliant even if the model’s code is, itself, on an open source license.

Coming back to the Deepseek model and seeing how well it performs until one digs deeper and finds it wanting. I am not sure if Deepseek will release the training data – I really hope they do. As of now, we can’t determine if the censoring of history is built into the data or somewhere else otherwise or if Deepseek is hallucinating.

Censorship of information in foundation models is something that we can run tests against. I would look at the Project Moonshot from the AI Verify Foundation as a starting point. We can run benchmarks and Red Teaming of these models using these recipes and extending it to include checks for historical information exclusion/removal.

In the grand scheme of things, foundational models and the training data need governance frameworks and there are already such frameworks already developed and being extended. Rewriting history is an important governance issue that we cannot shy away from – no matter what the political masters mandate.

As George Orwell astutely observed in his book 1984: “The past was erased, the erasure was forgotten, the lie became the truth.”