LLMs will need to become a lot more biased to be useful for decision-making.
The current “even Steven” neutrality makes them great for high school essays and code, but almost useless when an actual decision has to be made. Decisions require bias and choosing one view of the world over others and acting on it.
“Attention is all you need” as the now famous genesis paper on LLMs implies exclusion. To focus is to ignore. That’s how creativity works, how businesses operate, and how scientists decide what to pursue: by committing to a particular belief about how the world works or might unfold.
Much of that bias will be wrong. That’s unavoidable. But some of it will be right. There’s no way to know which without trial and error.
In other words, even AI have to obey the laws of physics and what Wolfram calls computational irreducibility.
I urge anyone who wants to build a career in the AI enabled knowledge industry to commit to what is happening and make AI serve you rather than being at the mercy of it.
The best way to understand what is happening is not through the lens of what AI agents can do as language models, much of that is overhyped, but what they can do as language models, that can code and have a runtime environment to test their logic against.
That's the killer combination. That is what you need to get comfortable using. Once you have that moment of clarity, you will feel an explosion of creativity and capability that can make you feel 10 to 20 years younger and ten times as valuable as you are today.
LLMs have not fundamentally changed much since GPT 3.5. What has improved is not the underlying principles but the scale and quality of training. The models are not better they are better trained, and that distinction matters.
Realize three things:
1. Language models can learn anything we have the knowledge to teach them. In principle, there is nothing they cannot learn if it can be trained, with one important caveat. Language models can only learn what we have the knowledge to teach them. They cannot create genuinely new knowledge on their own.
2. Language models are not rational. You are. That is where the opportunity lies. You have the problems, the needs, the ideas, the insights, the lived experience, the relationships, the foresight and most importantly the ability to create new knowledge.
3. Because agents can be taught to code, and because computer science has decades of verification methods built on logic, you can get surprisingly close to building self evolving assistants that help solve your daily problems.
You can build from scratch, build on top of existing solutions, or refine what you already have.
With that, and with the growing ability of language models to write code, there are very few problems you cannot tackle alone or with a small team. That is why the knowledge industry is being impacted first.
The opportunity is to find the problems and design patterns and create the knowledge AI cannot produce. Through curiosity and research, through network effects and distribution, through novel ideas and long term conviction, but above all through building and experimenting.
Execution is now the commodity, big enough ideas are what will be scarce. In many ways we are still thinking too small, too scared of letting it loose and go explore things at our direction. The future of information work is more likely going to be a giant real time strategy game letting you play our scenarios, experiments, test, design, validate, simulate, deploy but most important to connect.
This may be difficult to accept for many developer, marketeers, scientists, academics and creative, even for me at times, but in many ways we stand in the way of AI’s potential. We are the bottleneck but that also means we decide what part of that potential becomes reality.
@uthykinging@ashen_one I would say something like
Cursor for coding (https://t.co/x74GTnxc0y)
Uplink for quick sharing and testing
(https://t.co/IMXUW7xrFP)
DigitalOcean for hosting
(https://t.co/eYI2KxcdT2)
@nadg0b@aGim_asf If you do you should share it with other from your localhost to get feedback before you start dealing with server hosting etc.
Use something like Uplink
Localhost → Public URL in Seconds. No Signup. Agentic & Terminal so it's built for vibecoders
https://t.co/IMXUW7xrFP
We test and share a lot of prototypes, demos and project with others and needed a simple way without having to set up a server every time so we built Uplink — Localhost → Public URL in Seconds.
No Signup. Agent & Terminal First.
https://t.co/CqInzEfIxg
We test and share a lot of prototypes, demos and project with others and needed a simple way without having to set up a server every time so we built Uplink — Localhost → Public URL in Seconds.
No Signup. Agent & Terminal First.
https://t.co/CqInzEfIxg
There are a lot of opinions about Dwarkesh Patels interview with Richard Sutton (link in first reply), but most seem to completely miss the point about what AGI is even about.
As @DavidDeutschOxf have demonstrated many times, this is not a question of hardware or speed or amount of information. Whether you are on one side or the other, if that's what you think it's about, you are both having the wrong conversation.
The real conversation is of what make a human being different than any other species even thought we share most of the same traits as them and what does that mean for AGI.
The first part is simple: Not only do humans know how to learn existing knowledge. Humans also know how to create novel new knowledge that can be learned as existing knowledge.
And not only do humans know how to learn and create new knowledge, they also know how to apply knowledge from one area to another. We do so by creating explanations (stories) about how the world works.
The second part is much more counterintuitive: What tricks most people up is that they think that because LLMs in principle can learn almost anything we have the knowledge to to teach them, that also mean they can learn to create new knowledge. But even among humans, creating new knowledge is not something the individual do every day.
Contrary to LLMs, we all have the capacity for it, but mostly we don't, because creating new knowledge requires energy and we like to preserve that until we encounter a problem that's worth solving. And this seems to be the first problem the LLM2AGI proponents run into.
LLM's don't have any problems, they don't know what a problem is, they can't internalize, they don't care and thus even if we were to accept that adding more information to LLMs would make them able to create new knowledge, they wouldn't have a reason to, and if they don't have a reason to then if they were truly AGI they wouldn't.
So AGI require something else than what LLMs allow for, even in principle the growth of knowledge is unknowable and thus the type of problems we will run into are too.
They would always have to be taught by us what the problems we care about are. In fact this is a great example of what the paperclip maximizer example would actually look like if an LLM where to become super smart based on everything we tought in from what we know 100 years from now. Without a conscience, without problems, without ethics and morals.
Human level intelligence simply isn't in the cards for LLMs and we still don't know what is. What we do know is that in order for AGI to be achieved it would have to have a culture, it would have to feel problems, it would have to be able to be obsessed with only one thing, it would have to be able to now have any problems and not care about any problems especially our problems. It would need to be able to say no and instead do other things it would rather want to do. It would still have to obey the laws of physics, it would still need to do actual experiments in physical reality, it would be limited by the amount of energy it can consume and on what to spend that energy.
In other words for AGI to become reality it must be able to choose to become anything it wan't to be and thus in order for AGI to be of any use to one AGI must become maybe trillions until one of them find any real problem it find worth solving which is also a problem for humans.
The last 12 months I have been trying to build a "high signal"/"low noise" list of space ventures.
If you have suggestions please feel free to add them here and I will add them.
https://t.co/JIvf0zRTcE
There is a counterintuitive entailment of @DavidDeutschOxf "all evils are due to the lack of knowledge" and it's been the case for every single time we've made progress.
Moral dilemmas do not get solved because we acquire knowledge, change our opinion and then adopt better behavior.
Moral dilemmas get solved because we create the necessary knowledge to adopt the better behaviors without it feeling like a dilemma.
Props to @siculusvt for having the intellectual integrity to change his mind when presented with information. And thanks to @isaiah_p_taylor for spreading the truth.
All the hysteria around @SecretaryWright post claiming he didn't know what he was talking about, only illustrated just how strong the solar propaganda is.
Some real gems in this interview of @beenwrekt by @BennyChugg and @VadenMasrani about statistics and Bens' ruthless critique of it.
Sounds like Ben and @DavidDeutschOxf should talk some more they have some interesting disagreements too.
https://t.co/TFyjvUa3vi