Ex-Googlers take a stab at building 'general intelligence' that makes software do what you tell it Adept AI, an artificial intelligence R&D lab founded by ex-Googlers who helped invent the popular transformer architecture, launched on Tuesday with the ambitious goal of teaching machines how to use "every software tool and API in the world." The upstart raised $65 million from investors Saam Motamedi and LinkedIn co-founder …
"For example, you can use GPT-3 to talk about ordering a pizza, but you can't actually get it to do that for you,"
You had me at pizza.
But on a more serious note, $65m represents a serious investment and serious investments are often driven by business plans with revenue streams. Do we really need ex-Googlers trying to out Google Google? Will Dominos and Pizza Hut compete to get the company to use their API to fulfill my pizza craving?
I have a better idea: the pizza restaurants can have a web ordering tool that works. When they make it easy to order things quickly, they'll get customers. They don't need to use a massive language processing neural network in order to understand that they should make a pizza that's already on a menu and send it to an address. An address lookup field and optional cache of previously-used addresses and a menu that is easy to navigate costs a lot less. You can also easily modify that to support pick up or order ahead options as supported by the restaurant without having to re-teach the AI that they're available. There does not need to be any guesswork about whether your order worked.
"but if you don't have the ingredients you want"
The ingredients in pizza, being used in hundreds, if not thousands of other food stuffs, should be available in any kitchen.
"access to a very hot oven"
Adequate toaster ovens can be had for under twenty bucks at junk shops.
"or the desire to do the assembly"
I have no answer for sheer, bloody-minded laziness.
Laziness, really? Have you never been busy and eaten food cooked by someone else? Never used a restaurant?
Also, have you never used an ingredient and run out of it so you had to buy some more and couldn't make anything requiring it until you did? I tend to buy perishable things in quantities I expect will be fully used to avoid wasting it, which means that there are times where I used more than I predicted and have none left until I go to buy more.
And finally, have you ever been somewhere without a kitchen? Sometimes, when I'm not at home, I failed to take a toaster oven with me. I'm always forgetting it along with my laptop power cable.
"All of what you wrote indicates a distinct lack of planning."
No, it really doesn't. Take a situation where I am working late at an office on some important project for my employer. The office may not have an available oven, and I'm unlikely to carry a toaster oven with me on the odd chance I choose to use it. The office kitchen won't happen to include all the ingredients for pizza. If I tried to bring all those ingredients with me, some of them would be damaged or spoil. And, if I'm working late at an office, I likely don't have the time to fully make a pizza. Even if I knew all this would happen, there's not a lot I could have done. Most often, when something like this happens, I didn't know I'd have to do that until I was already there, preventing me from planning anything before coming into the office. I could of course run down to a grocer, buy the ingredients, and make a sort of pizza with the office microwave, but it will take longer and produce an inferior result than going to a restaurant and bringing one back.
Your argument works great if you're in your own house or the house of someone who doesn't mind you using their stuff. I'm guessing that, like me, you're in that situation a lot these days. Now that I'm working from home most of the time, I get the chance to cook many of my meals using my own equipment and ingredients, and I enjoy doing so. There are times where you're not at home, and things are different then. On occasion, some of those situations also apply at home. This office example is not the only one where that applies, but as you appear to have an investment in not accepting this, I'm not sure listing others would help.
Personally, I always keep something handy that'll allow me to get me through a few extra hours without getting hangry ... Again, it's hardly my fault if you are unprepared.
It's also hardly my fault if you're going to hold your breath until you are sick if you don't get your pizza when you demand it. Some of us don't sweat the petty stuff, life's too short.
"Adept will focus on training a neural network to perform general tasks on your computer, like generating compliance reports or plotting data using existing software such as Photoshop, Tableau, or Twilio, or designing humanoid-shaped killing machines." FTFY
Never underestimate the ability for mankind to destroy itself out of ignorant intelligence.
Search stack overflow for similar problem; copy-paste…
If someone can actually pull off what is suggested then it is evolutionary. Computer; create a model based on IEC standard xxxx. I don’t think it’s beyond wit to be possible to issue such a command. What I/O do you want to go with that? (assume flat text file in absence of spec?) choice of unit systems. Choice of language and modules?
Parse the standard for algorithms and data structure; and you are most of the way there.
The standard in particular I’m thinking of is a standard training exercise for our engineers to go away and create an implementation of it; and by far the best way to learn how it works. Python, excel, Java and fortran and others have all been used to different degrees of success. There isn’t a standard implementation included in the IEC but paper examples are given for calibration/testing purposes. Some commercial implementations have definite errors in of course too; which makes understanding at lower level very useful over trusting black boxes.
So they are trying to make the computer reason at a meta level and maybe meta-meta level. Just as Tesla discovered with early versions of Autopilot, you can make it see things at face value but unless you combine that with additional context from short and long term memory (ie what just happened and previous experience), with rules/laws etc the NN is not going to be able to drive flawlessly. And it isn't just a case of adding those contexts alongside what you have.
TMC tried to create these meta rules 25 years ago but got stuck in between rules, too-slow computers and the inefficient ML algorithms of the day.
Not sure I want chip sentient intelligence, droid slavery, droid armies and general unemployment starting with programmers.
I laughed at idiots who believed the drivel about 5GL programming languages and I still do. We are many, many, many years and iterations away from any form of ML actually generating efficient and decent code.
We're still suffering from idiotic mathematicians who only a few years tried to state that all computer programs can be described mathematically and therefore that 5GL would just generate the code. Their pathetic early examples generated some of the most stupid and inefficient and still very limited (barely more than "hello world") type code.
...any form of ML actually generating efficient and decent code...
The problem with machine generated code is that any programmer worth his or her salt will spend more time checking the generated code than they would have taken to write it themselves.
Any programmer using machine generated code without checking it first is tempting fate (and not in a good way).
No, the problem with machine generated code is the exact same problem as with human generated code. The hard part is getting the requirements right!
If you provide 100% bulletproof and fully complete specs, there are millions of programmers who can give you what you want. The reason you don't get what you want is because you don't know what you want, or in the rare cases you do you leave out a lot of detail you may "assume" but the person reading it doesn't have the context to come to the same assumption.
The hurdle to having a machine do that is it has to understand English (or whatever human language you write the specs in) including all the associated idioms, charts, graphs, etc. that may be included. If it requires the specs be in a certain machine readable format then you have to ask yourself, are they still "specs" or is that just a higher level computer language and a smarter compiler?
"every software tool and API in the world."
That's a problem, right there. It will spend all its time training as the rest of the software world adds new tools & APIs and makes breaking changes to existing ones.
"carry out someone's commands on its own initiative"
I'm not sure "someone's commands" and "its own initiative" aren't two separate and conflicting things.
Understanding what the user wants is only half the problem. Interfacing with the systems that make it possible to do what the user wants is a lot harder. You pretty much immediately have to implement a new API for every use case, and nothing scales. Entire industries need to adopt common standardized API for this to be successful.
But it's certainly enough for a VC pitch and possibly enough to get bought up for $X billions by $CORPORATION
"To us, it sounds like the computers in Star Trek, in which you describe an action to take, and the system figures out how to carry it out. It doesn't sound like too much of a stretch."
It does to me, because I've seen what GPT3, supposedly the better of the two and the ultimate in NLP text generators, prints when it doesn't have an answer to copy-paste. If whoever wrote that sentence doesn't know what I'm talking about, they should read the GPT articles in this very journal, because it's been covered a lot of times. GPT3 prints confusing, useless, and contradictory information all the time, and if we anthropomorphize it too much, it makes up junk whenever it doesn't know the answer. Now we expect something built the same way to understand how to interact with software that's designed for human users and probably has no programmatic interface. The best thing we can hope for is that it will at least figure out which keyboard shortcuts Microsoft has changed in Office, which would make it smarter than me, but actually carrying out instructions without flailing around is expecting too much.
"Luan envisions a future where people will be able to instruct the model to perform actions using natural language
Given the ambiguities inherent in human natural languages and the heavy human reliance on previous experience and context for understanding, I'm not exactly holding my breath until this works - except in very narrowly prescribed domains of vocabulary and action.
Burning through VC millions is hardly ludicrous. It keeps a roof overhead and the dawgs in kibble. Just like research grants at Uni. Many people are very good at making a living at this. In essence, it is a modern-day cottage industry.
I'm not here to discuss the rights and wrongs ... just pointing out facts.
"<i.[...] a modern-day cottage industry</i>
Cottage industry was, and in many cases still is, a subsistence necessity, so this is not really the best comparison. Money for old rope warpped in Emperors' clothes might be closer to the mark.
Which is why I said "In essence, it is a modern-day cottage industry.". See the modifiers? I placed them where I did for a reason.
To the folks burning the dollars, it's a subsistence necessity supporting their lifestyle.
... make my bed, bring my flying car around to the front, change the sprog's nappies, dig the spuds, pick the tomatoes and make & pressure can tomato sauce, mow the lawns (and take care of the equipment required), do the dishes and the laundry (including properly putting away both), fetch me the snail-mail and a beer, and the sprog's nappies will need changing again.
Come to think of it, I think I'll wait until the sprog is old enough to be trained to do the above. It'll be easier, faster, much, much cheaper and a lot more fool-proof.
Actually, the sprog is in her thirties ... and doing a pretty good job of training her own sprog to do all the above. Except the flying car, of course.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
