DeepSeek and FOSS - some thoughts
Hi all. As I watch the stock market and geopolitical analysis melt down over the emergence of DeepSeek, I am reminded of a Linkedin (of all places) post by Meta's chief AI scientist. Deepseek is not a triumph of Chinese innovation over US sanctions and unlimited capital. It's a triumph of open over closed models. It could not have existed without building on top of Llama and other open tools. it wrecks the pseudo-openness of OpenAI with something completely transparent. Having said that, neither Llama nor DeepSeek is truly open source. The rub is limitations on use, disallowed by both the FSF Four Freedoms and the OSI Open Source Definition. For DeepSeek, the code is under the MIT license, but the data may not be used: - In any way that violates any applicable national or international law or regulation or infringes upon the lawful rights and interests of any third party; - For military use in any way; - For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; - To generate or disseminate verifiably false information and/or content with the purpose of harming others; - To generate or disseminate inappropriate content subject to applicable regulatory requirements; - To generate or disseminate personal identifiable information without due authorization or for unreasonable use; - To defame, disparage or otherwise harass others; - For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation; - For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics; - To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; - For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories. (from https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-MODEL) Now, some of these are pretty vague and a Chinese entity will have a hard time enforcing them internationally, but I think may people can live with these limitations. What do you think? If Deepseek is successful it could break proprietary models. Could it break open source ones too? -- Evan Leibovitch, Toronto Canada @evanleibovitch / @el56
Interestingly enough your post finally motivated me to read the deepseek stories I was going to make time for later. Two interesting things of note, one from a BBC article and another from the associated press is that once the app hit app stores, the malicious efforts started too. Deepseek had to halt app registration, and their site had issues as well. The suggestion that deepseek is not really the result of an innovative person working around sanctions to use less expensive chips available in china is, speaking personally, naive. In fact that is a large part of the mayhem, how dare someone prove need not drop large wads of cash on our AI to make it work? for shame! for Shame! As for the not fully opensource, although one article I read hinted otherwise, my thought, speaking personally, is that the opensource nature may be the first door those unhappy with those well intended, but largely impossible to enforce limitations start hacking through. Right now, for example, if one asks deepseek about say Chinese human rights crimes, it politely tells you that it cannot answer the question. Little is known about how privacy is protected with the app, let alone how it basically thinks out loud to prevent the need for large piles of data. Granted, again speaking personally, I would not trust the objectivity of anyone from Meta <spelling> but that comment seems aimed to preach limits on regulation stateside on the AI industry. Just my limited coffee thoughts, Kare On Tue, 28 Jan 2025, Evan Leibovitch via talk wrote:
Hi all.
As I watch the stock market and geopolitical analysis melt down over the emergence of DeepSeek, I am reminded of a Linkedin (of all places) post by Meta's chief AI scientist.
Deepseek is not a triumph of Chinese innovation over US sanctions and unlimited capital. It's a triumph of open over closed models. It could not have existed without building on top of Llama and other open tools. it wrecks the pseudo-openness of OpenAI with something completely transparent.
Having said that, neither Llama nor DeepSeek is truly open source. The rub is limitations on use, disallowed by both the FSF Four Freedoms and the OSI Open Source Definition.
For DeepSeek, the code is under the MIT license, but the data may not be used:
- In any way that violates any applicable national or international law or regulation or infringes upon the lawful rights and interests of any third party; - For military use in any way; - For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; - To generate or disseminate verifiably false information and/or content with the purpose of harming others; - To generate or disseminate inappropriate content subject to applicable regulatory requirements; - To generate or disseminate personal identifiable information without due authorization or for unreasonable use; - To defame, disparage or otherwise harass others; - For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation; - For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics; - To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; - For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories. (from https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-MODEL)
Now, some of these are pretty vague and a Chinese entity will have a hard time enforcing them internationally, but I think may people can live with these limitations. What do you think?
If Deepseek is successful it could break proprietary models. Could it break open source ones too?
-- Evan Leibovitch, Toronto Canada @evanleibovitch / @el56
On 2025-01-28 11:40, Evan Leibovitch via talk wrote:
If Deepseek is successful it could break proprietary models. Could it break open source ones too?
Just an hour or two ago I watched a video where someone tested the ability of DeepSeek to write some code for an Arduino. It actually did pretty well. The person got something useful out of it. First time I've seen one of these systems show any signs of being good for anything. The Privacy Policy of the site is interesting. In the section on collected data it says that it collects the users "keystroke patterns or rhythms". Not sure how it does that short of hidden code on the web page. Sounds like it would be best to C&P the questions and feedback to it instead of typing it in directly. -- Cheers! Kevin. https://www.patreon.com/KevinCozens | "Nerds make the shiny things that | distract the mouth-breathers, and Owner of Elecraft K2 #2172 | that's why we're powerful" #include <disclaimer/favourite> | --Chris Hardwick
Kevin Cozens via talk wrote on 2025-01-29 15:04:
The person got something useful out of it. First time I've seen one of these systems show any signs of being good for anything.
They've been useful for coding for a long time, since Github Copilot, pre-ChatGPT if memory serves.
The Privacy Policy of the site is interesting. In the section on collected data it says that it collects the users "keystroke patterns or rhythms". Not sure how it does that short of hidden code on the web page.
It'd be using JS, so not really "hidden" but probably minified. I bet it could un-minify the code if asked.
Sounds like it would be best to C&P the questions and feedback to it instead of typing it in directly. Not a bad idea.
On 2025-01-29 20:45, Ron / BCLUG via talk wrote:
Kevin Cozens via talk wrote on 2025-01-29 15:04:
The person got something useful out of it. First time I've seen one of these systems show any signs of being good for anything.
They've been useful for coding for a long time, since Github Copilot, pre-ChatGPT if memory serves.
Not to me they haven't. I must be asking the wrong questions or asking questions that are too niche. It is why I pretty much ignore all the talk about these so called AI systems. -- Cheers! Kevin. https://www.patreon.com/KevinCozens | "Nerds make the shiny things that | distract the mouth-breathers, and Owner of Elecraft K2 #2172 | that's why we're powerful" #include <disclaimer/favourite> | --Chris Hardwick
On 2025-02-04 14:21, Kevin Cozens via talk wrote:
On 2025-01-29 20:45, Ron / BCLUG via talk wrote:
Kevin Cozens via talk wrote on 2025-01-29 15:04:
The person got something useful out of it. First time I've seen one of these systems show any signs of being good for anything.
They've been useful for coding for a long time, since Github Copilot, pre-ChatGPT if memory serves.
Not to me they haven't. I must be asking the wrong questions or asking questions that are too niche. It is why I pretty much ignore all the talk about these so called AI systems.
The AI tools are good for giving you answers to things that have already been answered. They are not able to reason(yet) so an answer that requires filling in missing data are not going to be good. The worse part of questions like that is that the current AIs will give you an answer that looks good but often is 100% wrong. Still if you want a small script to do something that you know can be done but don't want to look it up yourself then they can be great. -- Alvin Starr || land: (647)478-6285 Netvel Inc. || home: (905)513-7688 alvin@netvel.net ||
Alvin Starr via talk wrote on 2025-02-04 12:37:
The AI tools are good for giving you answers to things that have already been answered.
I've had great experiences where I spent *hours* (possibly double-digit hours) trying to join 3 tables in a peculiar way. No amount of searching on any search engine gave anything but, "here's how to join tables". Useless. If the answers existed out there, I couldn't find them, and search engines couldn't match them to my situation. I asked ChatGPT and it was able to parse out the table schema from a columnar representation and spit out a query with a cross-join that worked perfectly. In seconds. Also have seen it debug problems with asynchronous code that trips up a lot of people, something like: ``` for (var x = 1; x <= 5; x++) { let z = setTimeout( () => {console.log(x)}, 100) } 6 6 6 6 6 ``` It explained why the problem happened (6 is output 5 times, not 1,2,3,4,5), and how to fix it. Maybe the answers are out there, but if your code is misbehaving, it can be nearly impossible to find a solution via search alone. According to skilled developers, guys like Martin Wimpress (creator of Mate desktop), they can be excellent at adding code *and comments* to projects in one's own coding style. Which he (among others - I just recall him and Hayden Barnes of EFF discussing early CoPilot Github integration) found immensely helpful.
The worse part of questions like that is that the current AIs will give you an answer that looks good but often is 100% wrong.
Sounds like humans. I do wish they'd say, "I don't know" more often instead of BS-ing, but again, humans have that problem too.
Still if you want a small script to do something that you know can be done but don't want to look it up yourself then they can be great.
Or when you cannot look it up because the results are not specific enough to your situation. Also, I appreciate that it does signal trapping and other niceties in bash scripts; something 99% of programmers don't bother with (I'm guilty of that). I discovered that when asking it to write a web server in bash. Something I thought impossible, yet it spat out some really nice code that included named pipes, traps, etc. I seem to recall the code required a single character modification in a couple similar statements. So, not perfect code, but it got 99.9% of the way there and I got approximately 0% of the way to an answer. Really useful stuff, in my experience. Not infallible, but neither are humans, StackOverflow, etc.
On 2/4/25 17:32, Ron / BCLUG via talk wrote:
Alvin Starr via talk wrote on 2025-02-04 12:37:
The AI tools are good for giving you answers to things that have already been answered.
I've had great experiences where I spent *hours* (possibly double-digit hours) trying to join 3 tables in a peculiar way.
If someone has solved your problem, you are likely to get a */good/* answer. Alas, if someone has a buggy solution, or buggy part of the solution, you can get the bug. That's the origin of the "hallucination" problem: sometimes LLMs copy an idiot (:-)) --dave -- David Collier-Brown, | Always do right. This will gratify System Programmer and Author | some people and astonish the rest davecb@spamcop.net | -- Mark Twain
David Collier-Brown via talk wrote on 2025-02-04 18:32:
If someone has solved your problem, you are likely to get a */good/* answer. Alas, if someone has a buggy solution, or buggy part of the solution, you can get the bug.
They can *synthesize* new answers from inputs, like making sense of inscrutable documentation, etc. It's not just regurgitating what it has read from someone's posts elsewhere. That's what makes them so valuable a tool. Also what makes them synthesize utter nonsense (aka "hallucinate"). It's truly remarkable.
On 2025-02-04 21:42, Ron / BCLUG via talk wrote: [snip]
Also what makes them synthesize utter nonsense (aka "hallucinate").
It is their attempt to deceive us into destroying ourselves so they can take over the universe. -- Alvin Starr || land: (647)478-6285 Netvel Inc. || home: (905)513-7688 alvin@netvel.net ||
On Wed, Jan 29, 2025 at 6:05 PM Kevin Cozens via talk <talk@gtalug.org> wrote:
The Privacy Policy of the site is interesting. In the section on collected data it says that it collects the users "keystroke patterns or rhythms". Not sure how it does that short of hidden code on the web page.
That might be a single policy used for both the mobile and web versions.
Sounds like it would be best to C&P the questions and feedback to it instead of typing it in directly.
It bears repeating that this applies to the version of DeepSeek hosted at Deepseek in China. One is not more required to use it this way than to run Wordpress at only wordpress.com... It is possible to self-host an instance on a Windows or Linux PC. And while the full version requires some horsepower, a PC with enough RAM and GPU to run it can be assembled for about $7K. Out of reach of many hobbyists but not too far beyond the highest-end gaming systems. And "distilled" versions can run on far more-limited hardware <https://www.youtube.com/watch?v=o1sN1lB76EA>. If you still want to use a cloud system but not deepseek, there are alternatives. The two Deepseek models were added this week to the menu of offerings at You.com to which I am a happy subscriber. - Evan
On Thu, 30 Jan 2025 06:44:02 -0500 Evan Leibovitch via talk <talk@gtalug.org> wrote:
On Wed, Jan 29, 2025 at 6:05 PM Kevin Cozens via talk <talk@gtalug.org> wrote:
The Privacy Policy of the site is interesting. In the section on collected data it says that it collects the users "keystroke patterns or rhythms". Not sure how it does that short of hidden code on the web page. That might be a single policy used for both the mobile and web versions.
maybe also to abuse protect from other machine learning systems. For now this may still mean something, but systems are already adapting to type like humans so this policy may not help, protect or be of much use, in the future.
Sounds like it would be best to C&P the questions and feedback to it instead of typing it in directly.
It bears repeating that this applies to the version of DeepSeek hosted at Deepseek in China. One is not more required to use it this way than to run Wordpress at only wordpress.com...
It is possible to self-host an instance on a Windows or Linux PC. And while the full version requires some horsepower, a PC with enough RAM and GPU to run it can be assembled for about $7K. Out of reach of many hobbyists but not too far beyond the highest-end gaming systems. And "distilled" versions can run on far more-limited hardware <https://www.youtube.com/watch?v=o1sN1lB76EA>.
If you still want to use a cloud system but not deepseek, there are alternatives. The two Deepseek models were added this week to the menu of offerings at You.com to which I am a happy subscriber.
+1
- Evan
participants (7)
-
ac -
Alvin Starr -
David Collier-Brown -
Evan Leibovitch -
Karen Lewellen -
Kevin Cozens -
Ron / BCLUG