New AI Kid On The Block: Deepseek

[Timcognito]

If there's a computer program that's actually smarter than someone in practice, why employ that person? I've been asking this for decades now and never got a satisfying answer.

In many things in life there is acquired knowledge and hands on experience of success and failure , excepted best practice, trade secrets, black art and such that never makes it into or wrongly make it into publication and that is all that AI knows. Would you trust AI and robot to guide your surgery and a veteran surgeon. AI is just a high speed tool that is very beneficial to have in your toolbox to educate, save money and especially time but it needs someone to interoperate its findings, ask the detailed questions and those people will be in high demand. But AI makes nothing and those who do will also be in high demand.

 

[Somafunk]

From reading other tech forums the coding behind the modelling structure is written in C++/rust which is apparently many factors quicker and vastly more energy efficient than python of which the other ai’s use.

An economist that I follow, Stephanie Kelton asked Deepseek to create a model of what would happen to Canadian GDP if Trump imposed a 25% tariff on Canada.

Note x this is in no way an attempt to circumnavigate the “politics” ban but merely an example of just how detailed a response it can give

I studied Economics at “higher” level in the uk and the results and complete findings/modelling it returned in a few seconds are just mind boggling, if I handed in such a paper for the final exam I would’ve easily got an A-distinguished rather than my very stressful 3hr B mark

The Lens | Stephanie Kelton | Substack

Economic analysis, research, and commentary using the lens of MMT. Click to read The Lens, by Stephanie Kelton, a Substack publication with tens of thousands of subscribers.

stephaniekelton.substack.com

 

[amirm]

Don't ask him about Tiananmen…

Yeh, I saw the hardcoded answer to that in another video. The presenter argued that is no different than the hardcoded answers in other AIs with respect to our/western sensibilities.

 

[Gilvan]

The report from DeepSeek seems positive, except for the last section: "Criticisms" .
Considering that the sources of "criticisms" are just a bunch of deluded and "daydreaming" human jerks, I suspect that
AI will take the place of great part of humanity ...

 

[amirm]

I studied Economics at “higher” level in the uk and the results and complete findings/modelling it returned in a few seconds are just mind boggling, if I handed in such a paper for the final exam I would’ve easily got an A-distinguished rather than my very stressful 3hr B mark

The Lens | Stephanie Kelton | Substack

Economic analysis, research, and commentary using the lens of MMT. Click to read The Lens, by Stephanie Kelton, a Substack publication with tens of thousands of subscribers.

stephaniekelton.substack.com

Absolutely remarkable! So much more can be learned from reading its reasoning if not more than its final answer.

 

[kemmler3D]

I is just a high speed tool that is very beneficial to have in your toolbox to educate, save money and especially time but it needs someone to interoperate its findings, ask the detailed questions and those people will be in high demand. But AI makes nothing and those who do will also be in high demand.

This is a statement of AI's capabilities as they stand today. The point is, those limitations are defined as falling away with the advent of AGI, i.e. true human-level performance. You may believe that to be impossible, but there are many who think it will happen before the end of the decade or even the year.

Granted specialized knowledge and experience is worth something. Once AI acquires the useful knowledge that isn't published online or in digitized books, which could take... a few years, maybe? That closes another door for humans.

The way I see it, we've got:

1) Jobs where there's inherently little/no point in having a machine do it, like professional poker player or athlete.

2) Jobs where robotics capability hasn't caught up to people.

#2 only lasts until AGI develops better robots for us.

 

[amirm]

From reading other tech forums the coding behind the modelling structure is written in C++/rust which is apparently many factors quicker and vastly more energy efficient than python of which the other ai’s use.

I asked Deepseek this. It said OpenAI uses combination of Python, C++ and CUDA. I then asked it what it uses and it said the same including some of the same Python libraries. So the difference must be due to actual algorithm as opposed to programming language.

 

[JSmith]

Deepseek

Here's some comparison data;

https://artificialanalysis.ai/models/deepseek-v3

During training, the model learned to dynamically allocate more thinking time by reevaluating its initial problem-solving approach. This was the “aha” moment for both the model and the researchers! Rather than explicitly teaching the model on how to solve a problem, they simply provide it with the right incentives, and it autonomously develops advanced problem-solving strategies. A powerful reminder of the potential of RL to unlock new levels of intelligence in artificial systems

https://medium.com/@mayadakhatib/deepseek-r1-a-short-summary-73b6b8ced9cf

Looks like it's been attacked already too;

Chinese startup DeepSeek says it will temporarily limit registrations due to a cyber attack after the company’s artificial intelligence assistant amassed sudden popularity.

DeepSeek: Chinese AI startup hit by outage, cyber attack amid rising popularity

Chinese AI startup DeepSeek says it will temporarily limit registrations due to a cyber attack after it amassed sudden popularity.

www.smartcompany.com.au

The full training of DeepSeek-V3’s 671B parameters is claimed to have only taken 2.788M hours on NVidia H800 (Hopper-based) GPUs, which is almost a factor of ten less than others.

New Open Source DeepSeek V3 Language Model Making Waves

In the world of large language models (LLMs) there tend to be relatively few upsets ever since OpenAI barged onto the scene with its transformer-based GPT models a few years ago, yet now it seems t…

hackaday.com

JSmith

 

[RayDunzl]

I'll say. Nvidia is currently down almost 14% TODAY (including an after-hours improvement) which represents something like a drop of $400B in market cap. To put it in perspective, this is considered enough money to end world hunger for 10 years.

So, asking a presumedly American AI, it gives 733 million starving people, and at $40B/yr to "End World Humger", it comes out to $0.15/per day per starvee.


What's on the menu ? (I included snacks for extra calories)

Breakfast
Oatmeal with fresh fruits (bananas, berries) and a drizzle of honey
Whole grain toast with avocado spread
Boiled eggs or a vegetable omelet
Fresh juice or milk

Lunch
Grilled chicken or tofu with quinoa
Mixed green salad with tomatoes, cucumbers, and a light vinaigrette
Steamed vegetables (broccoli, carrots, bell peppers)
Fruit (apple, orange)

Snack
Yogurt with granola and a handful of nuts
Fresh fruit (grapes, apple slices)
Vegetable sticks (carrots, celery) with hummus

Dinner
Baked fish or lentil stew
Brown rice or whole grain pasta
Roasted vegetables (sweet potatoes, zucchini, bell peppers)
Side salad with a variety of greens

Midnight Snack
Whole grain crackers with cheese
Fruit smoothie (banana, spinach, almond milk)
Handful of nuts or seeds

This menu ensures a balance of proteins, carbohydrates, healthy fats, vitamins, and minerals. It also includes a variety of foods to keep meals interesting and enjoyable.

Uh-huh.


Where can I get even three meals like that for $0.15?

Finding three nutritious meals for $0.15 is extremely challenging given current food prices and economic conditions. The cost estimates I provided earlier are theoretical and based on large-scale, global efforts to end hunger, which would involve significant subsidies, donations, and efficient distribution systems.

Uh-huh.

In reality, affordable meal options can vary widely depending on location, availability of ingredients, and local economic conditions. For example, some budget-friendly meal ideas include:

Oatmeal with fruits and nuts
Rice and beans with vegetables
Pasta with tomato sauce and a side of steamed vegetables

These meals can be made relatively inexpensively, but even then, the cost would likely exceed $0.15 per meal.


So, the $40B is not enough, is it?

The $40 billion estimate is a theoretical figure that aims to address the immediate needs of those facing hunger. However, the actual cost could be higher due to various factors such as food prices, distribution challenges, and the need to address underlying issues like poverty, conflict, and climate change.


Food prices would be a problem?

Yes, food prices can be a significant challenge in addressing world hunger.


So, that $400B over ten years is Feel Good BS, isn't it?

The $400 billion estimate is a theoretical figure that aims to address the immediate needs of those facing hunger. While it is a significant amount, the actual cost could be higher due to various factors such as food prices, distribution challenges, and the need to address underlying issues like poverty, conflict, and climate change.

It said $400B over ten years, not "immediate" needs

You're right.

(some answer text omitted above, as AI likes to be very verbose, despite my protestations)

 

[Somafunk]

Absolutely remarkable! So much more can be learned from reading its reasoning if not more than its final

There’s a few repetitive phrases but the reasoning abilities are rather uncanny,

The way I see it, we've got:

1) Jobs where there's inherently little/no point in having a machine do it, like professional poker player or athlete.

2) Jobs where robotics capability hasn't caught up to people.

#2 only lasts until AGI develops better robots for us.

You forgot manual trades, could be good and raise the reward/pay level for such jobs/professions

 

[pablolie]

This is a statement of AI's capabilities as they stand today. The point is, those limitations are defined as falling away with the advent of AGI, i.e. true human-level performance. You may believe that to be impossible, but there are many who think it will happen before the end of the decade or even the year.

Granted specialized knowledge and experience is worth something. Once AI acquires the useful knowledge that isn't published online or in digitized books, which could take... a few years, maybe? That closes another door for humans.

The way I see it, we've got:

1) Jobs where there's inherently little/no point in having a machine do it, like professional poker player or athlete.

2) Jobs where robotics capability hasn't caught up to people.

#2 only lasts until AGI develops better robots for us.

AI doesn't innovate. Has zero ability. It is big scale plagiarizing of existing information - with zero ability to grow beyond it.

 

[amirm]

AI doesn't innovate. Has zero ability. It is big scale plagiarizing of existing information - with zero ability to grow beyond it.

The big question is: how much of our human endeavor is the same thing. It sometimes takes me hours to find an answer on the Internet for a problem I want to solve. If AI can provide that answer instantly, then that would be so many fewer hours someone would need to pay me to fix something.

Clearly some are overestimating AI abilities. But the other way around may also be true.

Sam Altman said a few days ago that every company he speaks to has plans to integrate AI. I know I like to do the same for ASR. So maybe AI becomes the same has having a website and support online for your company.

 

[amirm]

I think no matter what, we all need to increase our level of knowledge about AI and its implications for us. Lest we want to be left behind the technology trend. Hence this thread.

 

[kemmler3D]

AI doesn't innovate. Has zero ability. It is big scale plagiarizing of existing information - with zero ability to grow beyond it.

That's not a fundamental trait of AI, just how the most popular consumer-facing tech works today. And, it's not even true of AI in general. Some of the Google AI projects like AlphaFold and AlphaGo have introduced novelties to the world already. Even relatively simple genetic algorithms have been used to innovate in incremental ways for years now. And AI systems are already being used for basic scientific research in limited ways.

If you are betting on today's limitations keeping your job safe indefinitely, don't. That's my POV.

Trillions of dollars aren't being invested because this is the next frontier in productivity software and we'll be able to replace Excel and Salesforce, they're investing because they think they can "disrupt" all jobs at once.

 

[Music1969]

The fact that DeepSeek is a keystroke logger

I'm pretty sure they are all keystroke loggers.
I wouldn't even believe anyone's privacy policies.

 

[kemmler3D]

So, asking a presumedly American AI, it gives 733 million starving people, and at $40B/yr to "End World Humger", it comes out to $0.15/per day per starvee.


What's on the menu ? (I included snacks for extra calories)

Breakfast
Oatmeal with fresh fruits (bananas, berries) and a drizzle of honey
Whole grain toast with avocado spread
Boiled eggs or a vegetable omelet
Fresh juice or milk

Lunch
Grilled chicken or tofu with quinoa
Mixed green salad with tomatoes, cucumbers, and a light vinaigrette
Steamed vegetables (broccoli, carrots, bell peppers)
Fruit (apple, orange)

Snack
Yogurt with granola and a handful of nuts
Fresh fruit (grapes, apple slices)
Vegetable sticks (carrots, celery) with hummus

Dinner
Baked fish or lentil stew
Brown rice or whole grain pasta
Roasted vegetables (sweet potatoes, zucchini, bell peppers)
Side salad with a variety of greens

Midnight Snack
Whole grain crackers with cheese
Fruit smoothie (banana, spinach, almond milk)
Handful of nuts or seeds

This menu ensures a balance of proteins, carbohydrates, healthy fats, vitamins, and minerals. It also includes a variety of foods to keep meals interesting and enjoyable.

Uh-huh.


Where can I get even three meals like that for $0.15?

Finding three nutritious meals for $0.15 is extremely challenging given current food prices and economic conditions. The cost estimates I provided earlier are theoretical and based on large-scale, global efforts to end hunger, which would involve significant subsidies, donations, and efficient distribution systems.

Uh-huh.

In reality, affordable meal options can vary widely depending on location, availability of ingredients, and local economic conditions. For example, some budget-friendly meal ideas include:

Oatmeal with fruits and nuts
Rice and beans with vegetables
Pasta with tomato sauce and a side of steamed vegetables

These meals can be made relatively inexpensively, but even then, the cost would likely exceed $0.15 per meal.


So, the $40B is not enough, is it?

The $40 billion estimate is a theoretical figure that aims to address the immediate needs of those facing hunger. However, the actual cost could be higher due to various factors such as food prices, distribution challenges, and the need to address underlying issues like poverty, conflict, and climate change.


Food prices would be a problem?

Yes, food prices can be a significant challenge in addressing world hunger.


So, that $400B over ten years is Feel Good BS, isn't it?

The $400 billion estimate is a theoretical figure that aims to address the immediate needs of those facing hunger. While it is a significant amount, the actual cost could be higher due to various factors such as food prices, distribution challenges, and the need to address underlying issues like poverty, conflict, and climate change.

It said $400B over ten years, not "immediate" needs

You're right.

(some answer text omitted above, as AI likes to be very verbose, despite my protestations)

I don't know if the $40/400B number is accurate, it's just what comes up on Google, but I did want to emphasize that $400B is an absolute buttload of money.

 

[pablolie]

The big question is: how much of our human endeavor is the same thing. It sometimes takes me hours to find an answer on the Internet for a problem I want to solve. If AI can provide that answer instantly, then that would be so many fewer hours someone would need to pay me to fix something.

Clearly some are overestimating AI abilities. But the other way around may also be true.

Sam Altman said a few days ago that every company he speaks to has plans to integrate AI. I know I like to do the same for ASR. So maybe AI becomes the same has having a website and support online for your company.

Not in any way contradicting any of that.

AI most certainly powers more intelligent search functions. That (a) disrupts business models based on search and pitch, but for tech companies (like the one I work in) it (b) very powerfully allows people using complex products navigate day 0-1-2 scenarios with faster success... and those are over 60% of deployment cost according to many public studies.

The latter is the AI use case I work in - to make complex products easier to deploy.

There have been other evaluations of AI in ... creating code, fixing code and those sucked for now. I suspect that's an area AI can get much better in, but needs tons of training in, and with the vast variety of languages and code architectures we deal with it's not a task it's remotely been able to be very helpful in.

The most fundamental aspect is current AI models have zero ability to grow beyond initial and continuous human training, and none of the existing models indicate yet it ever will.

 

[pablolie]

If you are betting on today's limitations keeping your job safe indefinitely, don't. That's my POV.

If your job is training the model... yes, you are extremely safe. Especially if you take a basic AI certification and you do that. In fact you're at the center of what others perceive as "AI Innovation" then. AI entirely depends on humans to both create new "intelligence" as well as to train it into the model.

 

[Music1969]

It is not trained exclusively on the output of ChatGPT and other models

Understood, so it is training on outputs of other models and also other things.

In the paper they give the maths question example and it shows its thinking/reasoning.

But how do we know it hasn't already seen this same maths question and the 'working out' on the web somewhere.

I guess since it is open source and free for everyone to try, we can try to come up with unique questions to really test it.

I assume people are very impressed because it is giving very good solutions and thinking, to very unique questions?

 

[Petrushka]

I'll say. Nvidia is currently down almost 14% TODAY (including an after-hours improvement) which represents something like a drop of $400B in market cap. To put it in perspective, this is considered enough money to end world hunger for 10 years. TSMC down a similar percentage.

I had read somewhere that there were advances in computation that allowed model training to be done 95% faster / cheaper... I was wondering if that would affect Nvidia's business if it panned out. I guess now we know.

My AI expert says the advances are in software optimization. Some of it real improvements, and some shortcuts.

In any case, the software advances are not secret or proprietary.