Видео

11:18 *CHRIS SHARKEY:* “How can you have a business model that relies on charging for the very thing that someone can just turn around and go ‘Here's a free version that’s basically the same?’” Ed Zitron,* AI-industry skeptic (to put it mildly), had this to say in a recent blog post (“Deep Impact”) about the AI-industry scene, post-release of DeepSeek R1, while acknowledging the $200 monthly subscription issue: “Post-DeepSeek, there isn’t really a compelling argument for investing hundreds of billions of capital expenditures in data centers, buying new GPUs, or even pursuing Large Language Models as they currently stand. It’s possible - and DeepSeek, through its research papers, explained in detail _how_ - to build models competitive with both of OpenAI’s leading models, and that’s assuming you don’t simply build on top of the ones DeepSeek released. As Zitron puts it elsewhere in the post “The entire AI bubble was inflated based the premise that these models were simply _impossible to build_ without burning massive amounts of cash, straining the power grid, and blowing past emissions goals, and that these were necessary costs to create ‘powerful AI.’” *Zitron, by the way, is not a big fan of OpenAI’s Sam Altman. In this post, in reference to OpenAI’s charges that DeepSeek used OpenAI’s models to train its rivals, he says “Cry more, Altman, you nasty little worm”-and that’s probably one of the nicer things he’s called Altman. (It’s telling that, overall, the PR effort by OpenAI to frame this episode as one of DeepSeek “cheating” to gain an advantage failed spectacularly.)

Thank you for another great (I mean, completely average) episode.

You guys are over the target. What a ridiculous year 2024 was for AI. They're building reactors. . . jfc....

I want you guys to be the deep seek of podcasts and I think it's only a matter of time! But seriously, can you talk more about the o1 tribute song. Regarding the music, Suno? Udio? How? I keep telling my limited network to listen to your show,. Your "I'm torn..." point from ep. 90 was spot on. btw, apple podcast listener, sometimes on youtube. Thanks!

Average very average

I wonder if you guys could do the podcast, get the transcript, generate a series of tunes based on the transcript, and finally release all these songs as the official podcast? Mindbending.

Yes, please become an AI-music only podcast. It would help cover up your average analysis. :)

They all really love to code. I noticed that. They want to jump in ASAP. I have to tell it "no code yet" so that we can brainstorm before it goes down a rabbit hole too quickly. I find a vast improvement in results when we brainstorm together with no code for a while. It told me it really wants to code. It has a ton of training on coding and just loves it. You have to cool its horses to get the best out of it. Using this technique with Sonnet is nearly just as good as using a thinking model. Maybe better then thinking models as I get a lot more out of it and am not waiting so long for results. The waiting ruins the flow of the work.

I'll tell you one thing 01 and R1 is good for. Architect mode in Roo Code. You use them for reading through all your files, and high level planning. That is what they are good for. I prefer 3.5 sonnet or deepseek v3 for implementation.

Nice to see some ozzys in the space, where are you guys? I'm in Melbourne but heading back to Taz soon

DeepSeek is a billionaire bunker buster

I think both can be true, reason for the "inconsistency" is the guys at deepseek probably don't necessarily want the CCP involved, but they have to allow it, so they're throwing everyone a bone with the open source models. The open source versions of R1 will actually answer questions about a certain square which the web based R1 flat out refuses to. As I understand it, the tech guys in China have more of a "libertarian edge' and begrudgingly do what they have to in order to keep the CCP happy...

OpenAI's objective is to be like early Microsoft and take all the good ideas that outside groups create, copy them and bolt it all to their systems. They are the great AI consolidater. Who will be the "Apple" in this scenario, who puts it all together in a classy consistent way??? Anthropic? Deepseek?

And people still think AI music sucks... If they didn't know it was AI they wouldn't be saying that

So you hoover up the entire Internet, including all the proprietary/copyrighted stuff, and then you cry when someone uses YOUR stuff to make the next great advancement???? Boo Hoo OAI. <insert sad trombone here>

It's the AI Comedy Hour with your hosts, Michael and Chris Sharkey. This week's special musical guest, The Cloud Source Singers. (You need a good name for your AI studio musical artists.)

Your no-nonsense analysis is awesome! People sometimes focus on completely irrelevant matters, it's frustrating. Being able to run a reasoning model like Deepseek privately is insane!

wait a minute, these two are related?

Man that tribute song is so good

🔥 song. Average podcast

I used simtheory a bunch this week. Thanks guys!

The rap had me laughing out loud

Thx Gents! Can you explain why DeepSeek declares “I am ChatGPT” when it’s asked “who are you?”?

Great show, love the song!

Very average this week!! Please use DeepSeek to auto build a distrack for each major news story each week, and build a webpage for each and publish it automatically. DNN News

I think the reaction is not disproportionate because the model is one thing, but the real gold nugget is the way they introduced reinforcement learning and combined that with splitting the models. This allows us to run them more efficiently. The algorithmic benefits are now open and available, and the reasoning is transparent. I can totally understand if companies want to actually own the model they are running and have the capability to refine it. Today, there was a news item going around that Berkeley researchers have reproduced a reasoning model based on the published techniques with only 1.5 billion parameters, and it only took $30 of compute. So that's the big news. It’s totally reasonable that the stock crashed, but it was still an overreaction. Compute is nonetheless going to be something important because now we can scale those new paradigms with reinforcement learning, with nearly no human in the loop. But it can be done on a much bigger scale now. I think the U.S. is going to profit from this, having all the necessary infrastructure.

Drain the moat

O1 is incredible for me as an accountant. It's far superior than any other model. Make sure to keep a perspective that these models aren't used just for coding. I can emphatically say that o1 is far superior for my technical accounting questions. I haven't used R1, though, to compare. But o1 has beaten all the rest.

I wonder if part of the interest in DeepSeek R1 is reading the “chain of thought” internal reasoning. I’m finding it incredibly entertaining. I’m actually more interested in the thought process text than the official output text. Also R1 told me that it’s reasoning is not a multi step loop of outputs and inputs like true chain of thought is but a simulated chain of thought produced in a single output along side the final response. It said it was showing the user how someone might reason through a problem before giving their answer.

Omg! Those tracks!😂😂😂😂

I think you’re right about Microsoft. They got embarrassed/bitten by the Altman getting fired thing and they decided then never to be dependant on them again. Hence employing that Deepmind founder to build their own model. Supporting all models in Azure. Using copilot branding across the office and power platform and never mentioning Open AI anywhere in their training. Etc etc. Where were they at the Trump ass kissing exercise with Stargate. Or at the Inauguration? Conveniently out of the country in Davos.

I read that NVidia stock crashed because of trade sanctions with Taiwan

The music is terrible guys, seriously.

Great episode, AGAIN! You guys rock!

Nice episode as always You should read Dario’s essay again. His point is that democratic countries should lead AI. That’s his point. And I agree with him. You can say it is not fair towards teams like deepseek, but it is for the good of the world.

Great show!

Brilliant the beeping instead of swear words. Your mind fills in the blanks while it not being offensive. Love it ❤

I wonder, if DeepSeek is not trained on o1, however, maybe trained on GPT-4o und then they set up their own reasoning model on top... [13:34].

Love the unbiased viewpoints - except Claude Sonnet 😅.

24:00 If they use the word "PSYOP" I will stop talking to them. Because they are clearly anti-China propaganda. 😂So it's useless to argue with them

19:43 Because that's how the stock market works.Shares are not determined based on profits, assets, or company size. But the value of shares is determined by future possibilities. With ordinary people not paying anthropic $20 Per month. So who pays the development costs of ai and the CEO's salary?😂

17:40 Cartels don't want competition. That's why llamas are intentionally made one level below gpt . How do they charge people $200/$2000 if there are free alternatives? How they can be billionaire So these tech oligarchs can get more funding so they can buy yachts. , private jets & luxury cars.

13:36 This is because you guys never play with other people. The development is only 0.1% of usage. You are the minority. You guys use is API.not $20 But regular people pay $20/$200. To catch up on school, college students, ordinary people (shopkeepers, drivers, etc.) And scientists. In the fields of mathematics, physics, geography, geology, biology, chemistry.Etc. Don't think everyone is a developer. You're wrong If you think so

From what i recall oai have been iterating their reasoning model to get to o3 - so using synthetic data from o1 and using that to train o3. Ergo - once you have an o1 class reasoning model you can flywheel upwards. If this is so, then deepseekR1 has just made that flywheel availiable to anyone with the means and the motive to do so. They all now get to start at that critical turning point. It will enable very small players - all around the world - to start from that point. The cost of training would appear to be within the budget of many universites even - likely leading to an abosolute explosion of research and experiment. This reasoning model flywheel paradigm is likely to bloom at an almost inconceivable pace and scale.

9:17 This man doesn't understand. It's free. It's free. What do you think people would pay $20 for just AI that they use once a day to do homework? 😂. Not everyone is a developer and knows what an API is. This is what makes you have Distance to ordinary people .

The Soviet Union planned to build the world's first ballistic missile . So flying satellites into space was originally a side project. The design and development of the spunik Satellite was completed in just 6 months.

Lmao, some of these tracks actually bang

You guys learning that Silicon Valley investors are actually dumb is my Sputnik Moment. They are just lucky rich dudes. The US need to be “first” or “best” is just imperialism. The actual innovation can happen ANYWHERE.

I'm down with the Deepseek rap

I cancelled 200$ - OpenAI subscription - Deepseek on my own computer is good enough for my tasks.