As someone who has reported on AI for 7 years and covered China tech as well, I think the biggest lesson to be drawn from DeepSeek is the huge cracks it illustrates with the current dominant paradigm of AI development. A long thread. 1/
Comments
Log in with your Bluesky account to leave a comment
First, what is DeepSeek? A Chinese firm that was able to produce an open-source AI model with roughly 1/50th of the resources of state-of-the-art models yet still beat OpenAI’s o1 on several benchmarks. 2/
Much of the coverage has been focused on US-China tech competition. That misses a bigger story: DeepSeek has demonstrated that scaling up AI models relentlessly, a paradigm OpenAI introduced & champions, is not the only, and far from the best, way to develop AI. 3/
Thus far OpenAI & its peer scaling labs have sought to convince the public & policymakers that scaling is the best way to reach so-called AGI. This has always been more of an argument based in business than in science. 4/
There is empirical evidence that scaling AI models can lead to better performance. For businesses, such an approach lends itself to predictable quarterly planning cycles and offers a clear path for beating competition: Amass more chips. 5/
The problem is there are myriad huge negative externalities of taking this approach - not least of which is that you need to keep building massive data centers, which require the consumption of extraordinary amounts of resources. 6/
Many journalists have written extensively about those externalities.
When these massive data centers come to town, they distort power supplies (per @bloomberg.com’s @leonardonclt.bsky.social, @naurtorious.bsky.social,
@andretartar.bsky.social)… 7/
Curious why Apple’s basic on-device models and private cloud compute didn’t move the needle at all for months, but I suspect it’s down to just people’s wrong ideas about what Apple is good at.
What is the estimate of the resource usage of deepseek versus OpenAI? Wouldn’t the bake time for deepseek offset the energy usage and be a part of their business model… more for less as it were? Tech volatility could just be starting.
Sounds like an excuse for the ruling class to do what they want.
China showed them an efficient way of using resources. Say what you want about China but they aren’t invading countries (they generally invest), are investing in their own economy which we did in 1950s
AI capabilities were never held back by being too slow or not having enough resources, it was always held back by the inadequacy of the algorithms/heuristics/techniques.
That might be overstating things a little but it's a solid observation. I remember reading some AI executive talking to a reporter about how in the next year they'd have new chips online that would give them enough compute to solve the hallucination problem and realizing "this guy is clueless".
I like this thread a lot and I love that these terms are getting out there for everyone. This idea of chasing AGI has not been given enough light. It’s definitely a scary concept as AGI in business terms really means “capable of human work replacement”.
I'm also pretty sure you wouldn't get o1 or DeepSeek type of thinking before first scaling up the base model sufficiently. And you can still probably improve all that significantly by further scaling of the base model.
It's not one vs. the other but about the order needed for introducing things.
The OpenAI et al paradigm for scaling reminds me of how back in the 1920s & well into the '30s people tried to make more powerful engines by just bolting on superchargers. Sometimes multiple superchargers. Pretty quickly got to diminishing returns territory. ...
See, some engine designers had come to believe engines literally couldn't be made more powerful - that all future gains would come from supercharging.
They were wrong, of course. But whole product lines were based on this assumption.
Shocking how trying to do more by just throwing energy and resources at it until you will be fighting entropy was not the best option or solution. Almost like some effort should of been made toward optimization and efficiency?
I said this years ago and literally had massive push back and ostracizing for it. I left due to it. The brogrammer cult is extremely myopic and demanding of group think.
Deepseek is based on scaled up models from Meta and Alibaba that they fine tuned and applied reinforcement learning to to develop reasoning (Chain of Thought). This existing technique was applied in a clever but limited way. The true cost to develop this model is significantly understated.
The deeper problem here is that the US companies have promised improvements without top engineers, innovation is the opposite. The layoffs, RTOs, and churn due to moral have them falling behind.
Also though there's the question of what KIND of talent. Paying high salaries for people who know how to do basic stuff & assuming that if they just run those people in parallel, they'll get scaling. Basically, latest iteration of the Mythical Man-Month problem.
Is there proof that the DeepSeek computer is as good as its say it is? The idea that it created a competitive AI computer for @6M is very hard to believe. Could it be that DeepSeek, spent $6M but had funding sources such Chines government? When something looks too good to be true it usually is.
Can I ask if anyone has really evaluated these benchmarks? Looking at the DeepSeek repo, it's like 1300 lines around pytorch. If they're truly training on synthesized data, then I'd think this reveals holes in the benchmarks/tests more than anything (both suck, IMHO).
As an Information Security professional my opinion is China did this for $5 by stealing (hacking) everyone else, not the hard work of home-grown, ground up development.
There is nothing that could convince me otherwise.
Extra prevalent when dealing with infrastructure investment. Lots of African countries never built copper based internet infrastructure and now have mostly modern 4G/5G
This highlights for me what Sam is doing too. OpenAI has had “First Mover Advantage” so long as they have been able to maintain the claim that they are in a capex race. As soon as the tech isn’t shaped that way they no longer have an advantage from their position as first movers.
This was the issue Britain had vs. Germany during the 19th century. Britain had the initial jump, but then found itself already invested in older-generation capital equipment and hard-pressed to compete with Germany's later industrialization that could skip to the new gear without sunk costs.
Deepseek stood on the shoulders of Meta and Alibaba. Their model is a fine tuned version of Llama and Qwen with some RL enhancements. Without those LLMs it wouldn’t exist. It didn’t cost $5.6mm. 2,000 Nvidia A100’s cost $32mm. They didn’t release their training data and used multiple expert models.
The open source model they developed was based on Meta’s Llama 3.1 and 3.3 and Alibaba’s Qwen 2.5 all scaled up LLMs that cost hundreds of millions to develop and train. Comparing apples and oranges doesn’t describe the real story here.
A fascinating thread delving deep into the impact of artificial intelligence. There is definitely reason for concern about the safety and development, as well as its impact on society. The technology has its uses, but the effect it has on natural resources is something that should be considered.
It's clear serious actions need to be taken. I was always jealous of my Grandparents farm. Their water is glacier fed and pure. Ours on the other hand will turn you into an Oompa Loompa.
Incidentally and apropos nothing at all.
"The Hitchhiker's Guide to the Galaxy describes the Marketing Department of the Sirius Cybernetics Corporation as: 'A bunch of mindless jerks who'll be the first against the wall when the revolution comes.'"
Curiously, an edition of the Encyclopedia Galactica which fell through a rift in the time-space continuum from 1000 years in the future describes the Marketing Department of the Sirius Cybernetics Corporation as: A bunch of mindless jerks who were the first against the wall when the revolution came.
Love your post! So far the AI hype has exceeded its value. As an Engineer, it has some useful applications. For most others, it’s a toy or a way to write a fiction filled report in a fraction of the time it would take to actually research it and present facts.
Great thread. I think so many focus on the potential for AI to eliminate jobs or create some terminator future and ignore the immediate threat of the AI business itself, to the environment, to intellectual property, to society etc.
Great thread @karenhao.bsky.social ! Thank you for summarizing so well the terrible impacts of the current approach of big US AI co’s on health and humanity.
It is a brilliant thread. But my head hurts trying to understand the urgent rush to AI, unless it to save the planet from mankind. I'm 74 & remember when most homes didn't have a telephone or families a car. Each family had one bread winner, supported each other, gathered a meal times, kids played.
We learned in science how dynamos worked to 'ring forward, ring back' when connecting a call. You walked as often as possible to save pennies. Have we really progressed? Families lived in villages, cousins (I have 50 first ones) were your best reliable friends. You helped build each others houses.
You shared tools, nothing was thrown out, just recycled, food scraps fed to chooks, chooks gave eggs, then eventually a roast when too old. Cows milked & likewise Were we worse off? How is life really better? Do we need computers to do our thinking? We know how to save the planet, why don't we!
And the 'Conviviality of Commerce' has disappeared: Milton Friedman killed-off the 'social utility' [ie, conviviality] metric, so corps now use only 'shareholder value' to decide. And the glitter & power of Tech has helped kill-off Main Street, where commerce was fundamentally convivial
And the gen before you remember the advent of the telegram, the automobile and way before then, the printing press. This isn't even touching fire and the wheel.
Because exceptional individualism means each person gets to live in their Netflix Box and can avoid those pesky neoghbours. Or they can be digital nomads, surfing on beaches with only strangers around.
The Libertarian Dream is a goal for so many people. Bowling Alone details the US version.
You're a queen. I've been trying to articulate this exact thought for a year now, and haven't been nearly so clear. Will simply be pointing people to this thread.
Intellectual property theft?
OpenAI says it has evidence China’s DeepSeek used its model to train competitor | @financialtimes.com
"The San Francisco-based ChatGPT maker told the Financial Times it had seen some evidence of 'distillation', which it suspects to be from DeepSeek."
American AI firms try to poke holes in disruptive DeepSeek | @reuters.com
• Top US AI labs analyze DeepSeek's low-cost models
• Snowflake adds DeepSeek models amid customer demand
• DeepSeek likely spent more than widely reported $6 million figure, experts say
1/2 🧵
"DeepSeek's strides did not flow solely from a $6 million shoestring budget, a tiny sum compared to $250 billion analysts estimate big U.S. cloud companies will spend this year on AI infrastructure."
2/2 END🧵
Yes, massive AI datacenters could be a big problem, but:
• DeepSeek is 80% hype, 20% real.
• PRC is likely involved.
• China numbers are propaganda, not to be trusted.
• Deep skepticism warranted.
• Perplexity is real, impressive for accomplishing much with relatively little.
Yes, there is plenty of malarky on both sides.
The difference in the West is that there is also a great deal transparency.
The Gulf of America issue is just Trump flooding the zone with shit. It's what he does. Best to just ignore it. "Don't feed the trolls."
There is literally no transparency. OpenAI just went through a battle over this where the safety side lost. Most of the people involved on the core research and development have left.
Unlike OpenAI, DeepSeek released their model and a paper with great details about their training process so it's reproducible. As Yan Lecun said, this story appears to be about China vs America but it's really about open source vs closed source. We don't need to trust, we can verify.
OpenAI isn't the only alternative to DeepSeek,
and open source is not a panacea,
but there are several good open-source alternatives.
Some helpful resources:
1/4 🧵
What are Open Source Large Language Models? | IBM
Okay but China developing AI isn’t a bad thing. We need more countries involved. No one loves the Chinese government but the propaganda about them is ridiculous.
No, China developing AI is not a bad thing, the more the merrier, but China's military threats in the South China Sea and the Taiwan Strait are not propaganda—they are real and extremely dangerous.
A current active harm nobody in journalism seems yet to have noted: LLMs being granted authority to create the official judicial record (replacing actual human court reporters in the process). Flawed records can be tossed. Huge potential for corruption. This is happening all over the country.
You didn't mention degrowth at any moment in the discourse and it's probably one of the most degrowth threads I've ever seen.
Thank you very much for putting into a neat package every issue the AI has risen, in terms of consumism.
Is DeepSeek's efficiency breakthrough like AI’s Toyota Production System (TPS) moment? Just like Toyota turned resource constraints into lean manufacturing, DeepSeek shows that innovation thrives under limits. Is this the start of a lean AI revolution?
Arguably, OpenAI stopped innovating at all. They created a product that worked (got investors), then kept hectoring it to produce better output. But it has numerous, fundamentally unsolvable flaws.
DeepSeek’s success with less powerful chips shows how constraints fuel innovation. While the West poured billions into brute-force solutions, resource-limited teams in China built models that are leaner, cheaper, and just as powerful. Sometimes, too much money breeds inefficiency.
We had a similar experience at the end of the Cold War with numerical programming. The Russian and Eastern European programmers just got more done with less.
Engineering used to be about designing with constraints AKA "Any idiot can build a bridge that stands, but it takes an engineer to build a bridge that barely stands."
Throw unlimited budget at the problem and you get ... idiots.
Sounds like a re-run of why US cars became so poor until Japanese/European models forced the macho-power contest to look at efficiency &, you know, steering round corners and stuff.
So I guess US Tech can recover.
Just such a shame all those climate-killing AI boosters will have lost so much $$ 😀
Exactly, I had the same idea, its like the Toyota Production System (TPS) of AI moment? Just like Toyota turned resource constraints into lean manufacturing, DeepSeek shows that innovation thrives under limits. Is this the start of a lean AI revolution?
Not sure TechBros are creating consultative, ‘attention to detail’ businesses tho’! (Move fast etc)
One of my most stressful meetings was being quizzed by 4 Japanese engineers & technicians on a quality problem. The level of detail from these guys on a part in *someone else’s product* was amazing.
Oh, not our TechBros—they are narcissistic Icaruses.
They will probably take the DeepSeek shock moment and both ask their VC buddies and the government for a few additional billions.
Because it’s a moral fairy tale that gets used to justify resource allocation methods that are very harmful in fields of study and development that don’t have private VC cash firehoses
Not defending the AI companies. They can eat me, but the efficiency of the neglected underdog is just the flip side of the abundance/efficiency fallacy
Will be interesting to see the repsonse. Tons of money invested and cozying up to Trump for a reason maybe for a future block, win win for those involved right down to big oil and gas. Lose Lose for everyday folks that will pay the price literally and figuratively
What surprises me, is that some are so surprised. We had a pivotal somewhat unexpected moment with the introduction of attention and transformers pushing the capabilities of ML. It seemed unpredictable, but not unlikely, that "the next big thing" might pop up and eat the models that came before.
All these externalities are not the unintended consequences of an inconsiderate industry. They are the point. Capital is trying to heat the planet and use up our resources as fast as possible bc the greater the resource scarcity, the greater the emergency, the more power they have.
That's not the "model" responding that way, that's with filtering on the input prompt. If you download and use a local copy (the one I am quoting from is the 32B version), it responds with ...
No, it was my play on their slogan, “[American] Capitalism with Chinese characteristics”.
Instead of whine about hoe deepseek undercut American AI scale-up dynamics, then figure out how they did it first and then do it better without wasting money and natural resources.
Again, the model doesn't have these limitations. The limitations appear to be just in the public facing interface to it and filtering of the prompts, not the answers and not on the input data (for now).
Agreed…if this an LLM/chatbot-like program, it can train only on Chinese-gov’t-approved sources. But how much of that is baked into the program environment?
That I can’t answer. I’m not proficient in coding or AI. But if all of this is real, my hat is off to the DeepSeek folks on reducing waste.
The US kept stealth technology from falling into Russian hands so effectively the Russians had to develop the tech on its own, and largely did. China just laid the tracks for a railroad right next to the canal Sam Altman's built.
The Chinese were banned from buying huge piles of NVIDIA data center GPUs, so they engineered new, better methods to get the same result as US companies.
Railroads are better methods to move things vs canals. Canals are essentially rarely made anymore because of this.
One small point to add. DeepSeek used existing models to train. But that's not so important, as every AI model from now until the end of time will use everything every previous model learned.
I'm not an AI expert, but I think that's pretty much the entire point of AI. The problem is learning...
...on the fly. Someday they may learn to do this, but that day isn't any closer than it was when this AI hype began.
Personally, I don't think AI models will ever be able to do this, but I fear it doesn't matter. The more we turn to AI for answers, the more the inherent danger.
The fear of a Terminator 2 world misses the point. The real threat is we may find ourselves in world where facts are no longer the cognitive coin of the realm.
Reality is a subjective concept, at least to human minds. It depends on one's POV. AI may rob us of this cognitive flexibility. That...
I'd rather deal with the flexibility of human thought than an inflexible judgment of something that can never know what it is to be human. At least I can posit arguments to change the thinking of the former.
Fools love to worship stupidity. Idiots will start massive wars because inbred groups will believe their own Asinine.Incompetence. version that the sky is either green/red.
See #naziamerica for current post truth lunatic moronic collapse.
Great thread! I see parallels with chip design—the formulaic way to increase performance was continually reduce feature size and crank up clock speed, and it worked for a while. But if you let your capacity for other forms of innovation atrophy, you’ll lose in the end when other constraints intrude.
So ... seems the DeepSeek breakthrough is good in the sense that many environmental/industrial impacts of AI production may turn out to be more limited than thought.
Bad in the sense that if AI is cheaper/easier to produce with a smaller footprint, all of its *other* negative impacts will accelerate
College kids are writing essays, summarizing books they’re supposed yo read etc . Imagine that people in companies might be outsourcing job tasks too to AI
Yes, I get that summarizing of texts and writing essays are what students use it for. A horrible development, as people who can't summarize or write aren't learning much or developing important skills.
I'm wondering what kind of commercial tasks AI might be doing. Doing repetitive tasks; and ...?
Yes I think there’s a place for AI. I hear that coders find it enormously helpful in solving coding problems. Repetitive tasks like bagging items from major warehouses might be another good place to use. Though we should seriously slow consumption for sake of the environment…
Yes... which means Jevons paradox comes into play: a technology that consumes fewer resources per unit of output becomes so much more widely used that it consumes more resources in total
Something to bear in mind every time someone says a new, more efficient technology will help reduce emissions. Under capitalism, it probably won't and it might even do the opposite.
Thanks. I was about to say this. Now that these "reasoning" models are more accessible, there will be an increase in the number of companies attempting to maintain their own model, leading to a higher overall power consumption.
The scale and accessibility will be the problem still BUT at least it is not pushing all wealth to Silicon Valley and their pump and dump bullshit. Deepseek has shown the path to cheap Ai - the profit motive will change and hopefully for the better. So far it’s a brick
But I am hoping that these bloated companies sink to a point they’ll stop shoving this down our throats. No consumer asked for AI in fucking everything, only investors did.
"Artificial general intelligence (AGI) refers to the hypothetical intelligence of a machine that possesses the ability to understand or learn any intellectual task that a human being can. It is a type of artificial intelligence (AI) that aims to mimic the cognitive abilities of the human brain."
What's the value of AI, whether open source or Blockchain? What can it do for people, besides automate specific tasks, analyze data, code? How is it gonna put people out a work?
I ask bc last year I saw a brag that AI can write a grocery list, which ... doesn't seem worth all the resources it uses
I don't really know how AI is going to affect the workforce. I am a fan of ChatGPT and Microsoft's Copilot version of it. It excels at some things I need, like answering basic tech. questions. Ones like, how do I change a certain behavior in Windows, or how do I fix a problem I'm having in ...
Microsoft Outlook. I've also found it very good at finding tech. products by summarizing reviews. But while I think we're better for having Copilot available, I don't think the benefits they've advertised for integrating it into non-business Microsoft [Office] plans are too useful.
Here is Microsoft's help article on the benefits that Copilot adds to a Microsoft [Office] 365 Personal and Family plan. Not a compelling case for their price increase, in my opinion. I've found the free, standalone Copilot in Windows useful for some tasks but don't use for anything important.
Musk committed fraud to become citizen of USA. Strip him of his citizenship. Ban him from doing business in USA. Use RICO against him. Musk should spend many years in a state prison.
While you wait for the Trump Regime to act, consumers can boycott, investors can divest, and anarchists can vandalize Tesla. The vast majority of Musk’s wealth stands on that one pillar. Bring it down.
That hunt towards creating an AGI might be suicidal for humanity, at least according to the guy who was awarded a Nobel on the subject, because: "which intelligent entity/creature would accept taking orders from a less intelligent one?"
He gave the terminator scenario a 10-20% probability post-AGI
"agi" is a stupid, doomed and pointless goal, but definitely NOT for the reasons Hinton puts forward, which are just fanciful. Hinton is assisting the scammers with his rhetoric.
Innovation through necessity in a cornerstone of engineering/solutions. imho
We need to see who and how can push the ways we utilize the tools we have to the limits versus relying on compute power increasing. Then we utilize the latter better.
Seems to me it's a bit early to declare victory for DeepSeek. What happens when the Silicon Valley incumbents combine the optimizations that DeepSeek has pioneered with their access to much bigger compute clusters?
Play with it, it is worth it. There are a lot of negative opinions without much substance out there. People like to hate it, because it is easy, but that is shortsighted.
I'm not a coder, so all I know to do with it is ask questions. It's very fast at pulling info together, and summarizing pros and cons. I'm not sure the info is qualitatively better than I would get from an older search engine, but it can be presented in a more fulsome way. Other ways to play?
Also you can do so called few-shot prompting. I.e. you give it a few example tasks and their solutions and then let it solve the rest. Often this helps a lot to make it do the right thing.
And even for the search engine use case I find it helpful that you can ask follow up questions and clarifications to the search. I hope this helps, I am happy to discuss a more concrete example if you happen to have a specific problem at hand. It can for instance also answer legal questions.
In computer programming for example you can give it a problem and maybe some existing code and it will draft a solution. The more precise you are in posing the context and problem the closer the answer will be to something useful. It often makes all the difference.
A simple, but powerful example is you can let it translate texts. You can let it rephrase them for a specific target person. You can ask questions of how a specific person would think about what you are writing (which can also be used for marketing).
Effectively, a LLM is a trained simulator of all the speech on the internet, it can imitate and extrapolate the behaviours of any type of person it is has seen some amount of training data of.
Just imagine you would be on its end. You are an anonymous stranger that drops a random question/problem. If it knows more about you and the context it can answer more concretely. So in a sense you need to setup the narrative and be a bit of a machine whisperer to get it to do more specific things.
When a late stone age person noticed that every time he used green rocks around his fire, there'd be a blob of something weird in the cold ashes next day. Copper.
He/she didn't have the language or mental architecture to describe the bronze age, industrial revolution or the internet.
Ah -- kind of like we don't know what we don't know yet.
That makes sense ... the only things I can think of are already being done but more laboriously by humans. I don't conceptualize tasks we (I) haven't conceived of yet.
Eh, dude, no one is declaring any kind of “victory”: just generally marvelling and (justifiably) buzzing over a new, unexpected and potentially quite disruptive development in the field.
I'm not convinced of that. While there is currently no general solution for hallucinations, in some fields they can be mitigated (e.g. software engineering).
Even if hallucinations can be managed, a big if, this is a race to the bottom financially as none of the players have any real ideas for a competitive moat. LLM is commoditizing rapidly.
Maybe. I think it's too early to tell. We haven't seen many applications beyond chat bots yet, but they are coming (enabled by newer innovations like agents).
Imho AI will be back into the lab in a year or 2 with a shit tonne of broke investors, back where it belongs until an understandable and predictable core logic drives it- who knows when - 5 years, 50 years.
I've never heard anybody say or imply that "AGI solves Climate change" - but I suppose it's a reasonable statement.
What's a more reasonable statement is that decarbonization and mass-deployment of atmospheric carbon capture is ALREADY a solution to Climate Change. We knew this since the 1850's.
And frankly, given how much the price of renewables (esp Solar PV) has gone down, it's probably a whole lot fucking cheaper than chasing this AGI pie in the sky.
I think it's a reasonable assumption based on the basic physics of CO2 heat-absorption. Remove CO2, and the problem is solved (temporarily). (Sequestering the CO2 over geological time-periods is still pretty theoretical).
Doing this would be very much cheaper than what we're doing right now.
assumptions are easy, providing the enormous amounts of energy and materials required to actually do it is so hard that nobody is, except on trivial scale.
Great thread. I'm admittedly far from an expert on AI, but this more or less captures all of my various thoughts as I read today's news. Constraints drive innovation & it seems like the height of naivety for SV & policy-makers to view this as an arms race to be won by stockpiling & constraining.
Also worth exploring one of the most disturbing aspects of the pursuit of AGI - its eugenics roots. @timnitgebru.bsky.social and @xriskology.bsky.social are experts. https://youtu.be/P7XT4TWLzJw?si=EurRAKLKkAp18mCn
I’d also note that like every “high-on-their-own-supply” provocation of the MAGA administration, the Stargate hype was such a posturing provocation that it not only triggered President Musk, it also begged for a bitch slap from everyone in the tech world.
Not expecting China to call your bluff …
Comments
When these massive data centers come to town, they distort power supplies (per @bloomberg.com’s @leonardonclt.bsky.social, @naurtorious.bsky.social,
@andretartar.bsky.social)… 7/
https://www.bloomberg.com/graphics/2024-ai-power-home-appliances/
Deepseek might at least give us an extra month or 2 before we are obliterated!
In an endless loop of regression.
That's entirely ignored to perpetuate a scam for manipulation and masquerade for minority control that will kill our climate.
https://bsky.app/profile/karenhao.bsky.social/post/3lgq4pcb3532v
Very SV attitude. And a smooth segue from crypto mining. Not really "engineering", though.
But China apparently still has actual Engineers working the problem.
China showed them an efficient way of using resources. Say what you want about China but they aren’t invading countries (they generally invest), are investing in their own economy which we did in 1950s
Scale is always on top of architecture. Nobody treats it as "instead of architecture". This whole DeepSeek conversation feels like a straw man.
It's not one vs. the other but about the order needed for introducing things.
If your tech lead says "we just need more compute" you should hear "I don't have a solution, and I don't want to find one."
They were wrong, of course. But whole product lines were based on this assumption.
There is nothing that could convince me otherwise.
The real #HackingAI
If this principle doesn't have a name, it needs one.
In retrospect, that principle makes everything said about OpenAI and AI in general (trillion dollar industry, etc) even more ridiculous.
Thanks Karen! 👏👏
The fundamental design flaws were hidden by the superficial design flaws.
"The Hitchhiker's Guide to the Galaxy describes the Marketing Department of the Sirius Cybernetics Corporation as: 'A bunch of mindless jerks who'll be the first against the wall when the revolution comes.'"
The Libertarian Dream is a goal for so many people. Bowling Alone details the US version.
OpenAI says it has evidence China’s DeepSeek used its model to train competitor | @financialtimes.com
"The San Francisco-based ChatGPT maker told the Financial Times it had seen some evidence of 'distillation', which it suspects to be from DeepSeek."
• Top US AI labs analyze DeepSeek's low-cost models
• Snowflake adds DeepSeek models amid customer demand
• DeepSeek likely spent more than widely reported $6 million figure, experts say
1/2 🧵
2/2 END🧵
• DeepSeek is 80% hype, 20% real.
• PRC is likely involved.
• China numbers are propaganda, not to be trusted.
• Deep skepticism warranted.
• Perplexity is real, impressive for accomplishing much with relatively little.
The difference in the West is that there is also a great deal transparency.
The Gulf of America issue is just Trump flooding the zone with shit. It's what he does. Best to just ignore it. "Don't feed the trolls."
US just announced 500B of funding for OpenAI. The heads of several tech companies are infesting the government.
China numbers are propaganda? Unlike the extremely unbiased tech companies?
You've missed the forest for the trees.
and open source is not a panacea,
but there are several good open-source alternatives.
Some helpful resources:
1/4 🧵
What are Open Source Large Language Models? | IBM
Best Open Source LLMs of 2024 — Klu
Top 5 Free Open Source LLMs in 2024 | Cody
8 Top Open-Source LLMs for 2024 and Their Uses | DataCamp
Thank you very much for putting into a neat package every issue the AI has risen, in terms of consumism.
#AI #Efficiency #Innovation #Toyota
If we're not coding or analyzing market data, what do you think it's good for?
Engineering used to be about designing with constraints AKA "Any idiot can build a bridge that stands, but it takes an engineer to build a bridge that barely stands."
Throw unlimited budget at the problem and you get ... idiots.
So I guess US Tech can recover.
Just such a shame all those climate-killing AI boosters will have lost so much $$ 😀
One of my most stressful meetings was being quizzed by 4 Japanese engineers & technicians on a quality problem. The level of detail from these guys on a part in *someone else’s product* was amazing.
They will probably take the DeepSeek shock moment and both ask their VC buddies and the government for a few additional billions.
I'm not sure how you interpreted this as a generalization arguing for defunding all research.
AI art / writing aren't great use cases at all, and they're also utilized mainly by business.
Is there a reason US taxpayers should pour money into it that goes into billionaire pockets?
Plus DeepSeek is open source, meaning anyone can train their own AI energy efficient model. https://github.com/deepseek-ai
I am curious how models get their alignment to specific cultural norms, though. Is it a discrete step?
Look at it this way: Chinese AI with American characteristics.
If DeepSeek is really all that, then it’s a Sputnik moment.
Instead of whine about hoe deepseek undercut American AI scale-up dynamics, then figure out how they did it first and then do it better without wasting money and natural resources.
That I can’t answer. I’m not proficient in coding or AI. But if all of this is real, my hat is off to the DeepSeek folks on reducing waste.
https://www.secretprojects.co.uk/threads/soviet-origins-of-stealth.38185/
Railroads are better methods to move things vs canals. Canals are essentially rarely made anymore because of this.
I'm not an AI expert, but I think that's pretty much the entire point of AI. The problem is learning...
This is why even the most advanced AI models are far less than dependable if accurate information is the ultimate goal.
It's reassuring, in a way. They can't do what we do...
Personally, I don't think AI models will ever be able to do this, but I fear it doesn't matter. The more we turn to AI for answers, the more the inherent danger.
Reality is a subjective concept, at least to human minds. It depends on one's POV. AI may rob us of this cognitive flexibility. That...
I'd rather deal with the flexibility of human thought than an inflexible judgment of something that can never know what it is to be human. At least I can posit arguments to change the thinking of the former.
But hey, what the fuck do I know.
See #naziamerica for current post truth lunatic moronic collapse.
Powered by @skywriter.blue
Bad in the sense that if AI is cheaper/easier to produce with a smaller footprint, all of its *other* negative impacts will accelerate
I'm wondering what kind of commercial tasks AI might be doing. Doing repetitive tasks; and ...?
I ask bc last year I saw a brag that AI can write a grocery list, which ... doesn't seem worth all the resources it uses
You can paste the link
https://bsky.app/profile/karenhao.bsky.social/post/3lgq4pmrwgl2v
here
https://skyview.social/
https://www.newsweek.com/elon-musk-immigration-status-1990s-donald-trump-twitter-2024-election-1978588
https://www.muskwatch.com/p/how-do-you-explain-trumps-abrupt?utm_campaign=email-half-post&r=f879&utm_source=substack&utm_medium=email
He gave the terminator scenario a 10-20% probability post-AGI
https://bsky.app/profile/rahll.bsky.social/post/3leb4n5z4ic2q
We need to see who and how can push the ways we utilize the tools we have to the limits versus relying on compute power increasing. Then we utilize the latter better.
Yes, let's of money spent, data centers built, power consumed.
it's not a great analogy, but try this.
He/she didn't have the language or mental architecture to describe the bronze age, industrial revolution or the internet.
That makes sense ... the only things I can think of are already being done but more laboriously by humans. I don't conceptualize tasks we (I) haven't conceived of yet.
Thanks, that helps!
AI is an unavoidable phase. It's going to happen. The uses are mostly obscure or hard to grasp.
But it will change everything, fir good ir ill. And it's in process and won't stop to explain itself.
What's a more reasonable statement is that decarbonization and mass-deployment of atmospheric carbon capture is ALREADY a solution to Climate Change. We knew this since the 1850's.
We can't save the forest by burning down the forest.
Some capitalists are making bank promising it, but thats NOT the same thing.
Doing this would be very much cheaper than what we're doing right now.
Powered by @skywriter.blue
Not expecting China to call your bluff …
https://bsky.app/profile/techmeme.com/post/3lguaiyp5222u