Locklin on science

Andreessen-Horowitz craps on “AI” startups from a great height

Posted in investments by Scott Locklin on February 21, 2020

Andreessen-Horowitz has always been the most levelheaded of the major current year VC firms. While other firms were levering up on “cleantech” and nonsensical biotech startups that violate physical law, they quietly continued to invest in sane companies (also hot garbage bugman products like soylent).  I assume they actually listen to people on the front lines, rather than what their VC pals are telling them. Maybe they’re just smarter than everyone else; definitely more independent minded. Their recent review on how “AI” differs from software company investments is absolutely brutal. I am pretty sure most people didn’t get the point, so I’ll quote it emphasizing the important bits.

https://a16z.com/2020/02/16/the-new-business-of-ai-and-how-its-different-from-traditional-software/

They use all the buzzwords (my personal bete-noir; the term “AI” when they mean “machine learning”), but they’ve finally publicly noticed certain things which are abundantly obvious to anyone who works in the field. For example, gross margins are low for deep learning startups that use “cloud” compute. Mostly because they use cloud compute.

 

Gross Margins, Part 1: Cloud infrastructure is a substantial – and sometimes hidden – cost for AI companies 🏭

In the old days of on-premise software, delivering a product meant stamping out and shipping physical media – the cost of running the software, whether on servers or desktops, was borne by the buyer. Today, with the dominance of SaaS, that cost has been pushed back to the vendor. Most software companies pay big AWS or Azure bills every month – the more demanding the software, the higher the bill.

AI, it turns out, is pretty demanding:

  • Training a single AI model can cost hundreds of thousands of dollars (or more) in compute resources. While it’s tempting to treat this as a one-time cost, retraining is increasingly recognized as an ongoing cost, since the data that feeds AI models tends to change over time (a phenomenon known as “data drift”).
  • Model inference (the process of generating predictions in production) is also more computationally complex than operating traditional software. Executing a long series of matrix multiplications just requires more math than, for example, reading from a database.
  • AI applications are more likely than traditional software to operate on rich media like images, audio, or video. These types of data consume higher than usual storage resources, are expensive to process, and often suffer from region of interest issues – an application may need to process a large file to find a small, relevant snippet.
  • We’ve had AI companies tell us that cloud operations can be more complex and costly than traditional approaches, particularly because there aren’t good tools to scale AI models globally. As a result, some AI companies have to routinely transfer trained models across cloud regions – racking up big ingress and egress costs – to improve reliability, latency, and compliance.

Taken together, these forces contribute to the 25% or more of revenue that AI companies often spend on cloud resources. In extreme cases, startups tackling particularly complex tasks have actually found manual data processing cheaper than executing a trained model.

This is something which is true of pretty much all machine learning with heavy compute and data problems. The pricing structure of “cloud” bullshit is designed to extract maximum blood from people with heavy data or compute requirements. Cloud companies would prefer to sell the time on a piece of hardware to 5 or 10 customers. If you’re lucky enough to have a startup that runs on a few million rows worth of data and a GBM or Random Forest, it’s probably not true at all, but precious few startups are so lucky. Those who use the latest DL woo on the huge data sets they require will have huge compute bills unless they buy their own hardware. For reasons that make no sense to me, most of them don’t buy hardware.

In many problem domains, exponentially more processing and data are needed to get incrementally more accuracy. This means – as we’ve noted before – that model complexity is growing at an incredible rate, and it’s unlikely processors will be able to keep up. Moore’s Law is not enough. (For example, the compute resources required to train state-of-the-art AI models has grown over 300,000x since 2012, while the transistor count of NVIDIA GPUs has grown only ~4x!) Distributed computing is a compelling solution to this problem, but it primarily addresses speed – not cost.

Beyond what they’re saying about the size of Deep Learning models which is doubtless true for interesting new results, admitting that the computational power of GPU chips hasn’t exactly been growing apace is something rarely heard (though more often lately). Everyone thinks Moore’s law will save us. NVIDIA actually does have obvious performance improvements that could be made, but the scale of things is such that the only way to grow significantly bigger models is by lining up more GPUs. Doing this in a “cloud” you’re renting from a profit making company is financial suicide.

 

Gross Margins, Part 2: Many AI applications rely on “humans in the loop” to function at a high level of accuracy 👷

Human-in-the-loop systems take two forms, both of which contribute to lower gross margins for many AI startups.

First: training most of today’s state-of-the-art AI models involves the manual cleaning and labeling of large datasets. This process is laborious, expensive, and among the biggest barriers to more widespread adoption of AI. Plus, as we discussed above, training doesn’t end once a model is deployed. To maintain accuracy, new training data needs to be continually captured, labeled, and fed back into the system. Although techniques like drift detection and active learning can reduce the burden, anecdotal data shows that many companies spend up to 10-15% of revenue on this process – usually not counting core engineering resources – and suggests ongoing development work exceeds typical bug fixes and feature additions.

Second: for many tasks, especially those requiring greater cognitive reasoning, humans are often plugged into AI systems in real time. Social media companies, for example, employ thousands of human reviewers to augment AI-based moderation systems. Many autonomous vehicle systems include remote human operators, and most AI-based medical devices interface with physicians as joint decision makers. More and more startups are adopting this approach as the capabilities of modern AI systems are becoming better understood. A number of AI companies that planned to sell pure software products are increasingly bringing a services capability in-house and booking the associated costs.

Everyone in the business knows about this. If you’re working with interesting models, even assuming the presence of infinite accurately labeled training data, the “human in the loop” problem doesn’t ever completely go away. A machine learning model is generally “man amplified.” If you need someone (or, more likely, several someone’s) making a half million bucks a year to keep your neural net producing reasonable results, you might reconsider your choices. If the thing makes human level decisions a few hundred times a year, it might be easier and cheaper for humans to make those decisions manually, using a better user interface. Better user interfaces are sorely underappreciated. Have a look at Labview, Delphi or Palantir’s offerings for examples of highly productive user interfaces.

 

 Since the range of possible input values is so large, each new customer deployment is likely to generate data that has never been seen before. Even customers that appear similar – two auto manufacturers doing defect detection, for example – may require substantially different training data, due to something as simple as the placement of video cameras on their assembly lines.

 

Software which solves a business problem generally scales to new customers. You do some database back end grunt work, plug it in, and you’re done.  Sometimes you have to adjust processes to fit the accepted uses of the software; or spend absurd amounts of labor adjusting the software to work with your business processes: SAP is notorious for this. Such cycles are hugely time and labor consuming. Obviously they must be worth it at least some of the time. But while SAP is notorious (to the point of causing bankruptcy in otherwise healthy companies), most people haven’t figured out that ML oriented processes almost never scale like a simpler application would. You will be confronted with the same problem as using SAP; there is a ton of work done up front; all of it custom. I’ll go out on a limb and assert that most of the up front data pipelining and organizational changes which allow for it are probably more valuable than the actual machine learning piece.

 

In the AI world, technical differentiation is harder to achieve. New model architectures are being developed mostly in open, academic settings. Reference implementations (pre-trained models) are available from open-source libraries, and model parameters can be optimized automatically. Data is the core of an AI system, but it’s often owned by customers, in the public domain, or over time becomes a commodity.

That’s right; that’s why a lone wolf like me, or a small team can do as good or better a job than some firm with 100x the head count and 100m in VC backing. I know what the strengths and weaknesses of the latest woo is. Worse than that: I know that, from a business perspective, something dumb like Naive Bayes or a linear model might solve the customer’s problem just as well as the latest gigawatt neural net atrocity. The VC backed startup might be betting on their “special tool” as its moaty IP. A few percent difference on a ROC curve won’t matter if the data is hand wavey and not really labeled properly, which describes most data you’ll encounter in the wild. ML is undeniably useful, but it is extremely rare that a startup have “special sauce” that works 10x or 100x better than somthing you could fork in a git repo. People won’t pay a premium over in-house ad-hoc data science solutions unless it represents truly game changing results. The technology could impress the shit out of everyone else, but if it’s only getting 5% better MAPE (or whatever); it’s irrelevant. A lot of “AI” doesn’t really work better than a histogram via “group by” query. Throwing complexity at it won’t make it better: sometimes there’s no data in your data.

 

Some good bullet points for would be “AI” technologists:

Eliminate model complexity as much as possible. We’ve seen a massive difference in COGS between startups that train a unique model per customer versus those that are able to share a single model (or set of models) among all customers….

Nice to be able to do, but super rare. If you’ve found a problem like this, you better hope you have a special, moaty solution, or a unique data set which makes it possible.

Choose problem domains carefully – and often narrowly – to reduce data complexityAutomating human labor is a fundamentally hard thing to do. Many companies are finding that the minimum viable task for AI models is narrower than they expected.  Rather than offering general text suggestions, for instance, some teams have found success offering short suggestions in email or job postings. Companies working in the CRM space have found highly valuable niches for AI based just around updating records. There is a large class of problems, like these, that are hard for humans to perform but relatively easy for AI. They tend to involve high-scale, low-complexity tasks, such as moderation, data entry/coding, transcription, etc.

This is a huge admission of “AI” failure. All the sugar plum fairy bullshit about “AI replacing jobs” evaporates in the puff of pixie dust it always was. Really, they’re talking about cheap overseas labor when lizard man fixers like Yang regurgitate the “AI coming for your jobs” meme; AI actually stands for “Alien (or) Immigrant” in this context. Yes they do hold out the possibility of ML being used in some limited domains; I agree, but the hockey stick required for VC backing, and the army of Ph.D.s required to make it work doesn’t really mix well with those limited domains, which have a limited market.

Embrace services. There are huge opportunities to meet the market where it stands. That may mean offering a full-stack translation service rather than translation software or running a taxi service rather than selling self-driving cars.

In other words; you probably can’t build a brain in a can that can solve all kinds of problems: you’re probably going to be a consulting and services company. In case you aren’t familiar with valuations math: services companies are worth something like 2x yearly revenue; where software and “technology” companies are worth 10-20x revenue. That’s why the wework weasel kept trying to position his pyramid scheme as a software company. The implications here are huge: “AI” raises done by A16z and people who think like them are going to be at much lower valuations. If it weren’t clear enough by now, they said it again:

To summarize: most AI systems today aren’t quite software, in the traditional sense. And AI businesses, as a result, don’t look exactly like software businesses. They involve ongoing human support and material variable costs. They often don’t scale quite as easily as we’d like. And strong defensibility – critical to the “build once / sell many times” software model – doesn’t seem to come for free.

These traits make AI feel, to an extent, like a services business. Put another way: you can replace the services firm, but you can’t (completely) replace the services.

I’ll say it again since they did: services companies are not valued like software businesses are. VCs love software businesses; work hard up front to solve a problem, print money forever. That’s why they get the 10-20x revenues valuations. Services companies? Why would you invest in a services company? Their growth is inherently constrained by labor costs and weird addressable market issues.

This isn’t exactly an announcement of a new “AI winter,” but it’s autumn and the winter is coming for startups who claim to be offering world beating “AI” solutions. The promise of “AI” has always been to replace human labor and increase human power over nature. People who actually think ML is “AI” think the machine will just teach itself somehow; no humans needed. Yet, that’s not the financial or physical reality. The reality is, there are interesting models which can be applied to business problems by armies of well trained DBAs, data engineers, statisticians and technicians. These sorts of things are often best grown inside a large existing company to increase productivity. If the company is sclerotic, it can hire outside consultants, just as they’ve always done. A16z’s portfolio reflects this. Putting aside their autonomous vehicle bets (which look like they don’t have a large “AI” component to them), and some health tech bets that have at least linear regression tier data science, I can only identify only two overtly data science related startup they’ve funded. They’re vastly more long crypto currency and blockchain than “AI.” Despite having said otherwise, their money says “AI” companies don’t look so hot.

My TLDR summary:

  1.  Deep learning costs a lot in compute, for marginal payoffs
  2. Machine learning startups generally have no moat or meaningful special sauce
  3. Machine learning startups are mostly services businesses, not software businesses
  4. Machine learning will be most productive inside large organizations that have data and process inefficiencies

 

 

Shitty future: Bugman design versus eternal design

Posted in Design by Scott Locklin on February 4, 2020

I was yacking with nerds recently on the reason why some people enjoy owning  mechanical wristwatches. In the finance business or any enterprise sales org, wearing a mechanical wristwatch is well understood, like wearing a nice pair of leather shoes or a silk necktie. Tastes may differ, but people in that milieu understand the appeal. In tech, other than a small subculture  of people who wear the Speedmaster moon watch (because we all wanted to be astronauts), and an even smaller subculture who wear something like the Rolex Milgauss (some of us work around big atom-smashing magnets), the mechanical wristwatch is mostly a source of confusion.

You can dismiss it as an expensive status symbol (many things are; nice cars, nice bags, nice nerd dildo, nice anything), but the continued existence of the mechanical wristwatch is more than that. The wristwatch became popular after WW-1, and was a necessary piece of equipment in the time of the last great explorers, from the Everest and Polar expeditions to Jaques Cousteau‘s undersea adventures to the Moon landing. The association with this now historic, but still golden era continues to sell wristwatches.

The geared mechanical clockwork itself is ancient: we have no idea where/when it was invented, but we know the ancient Greeks had such mechanisms. While there is no evidence for or against it, it is possible that gear trains predate recorded civilization. The geared mechanical clock, like the pipe organ and the Gothic cathedral is a defining symbol of Western Civilization. Division of the day into mechanically measured hours  unrelated to the movements of the sun is a symbol of the defeat of the tyranny of nature by human ingenuity and machine culture.

As a piece of technology, wristwatches probably peaked around 1970 when quartz watches became a thing. Quartz watches are undoubtedly more accurate, and at this point you could probably stick a microdot which syncs to GPS atomic clocks anywhere. But the psychological framework, and the association with the last  human earthbound age of adventure and exploration remains. Watchmakers continue to innovate; my daily beater by Damasko contains a bunch of technology you usually only see in an experimental physics Ph.D. thesis (saw them all in mine anyway); ceramic bearings, martensitic steel, preferentially etched silicon springs, viton o-rings. None of this is necessary to build a good watch; it is just a tribute to the art of mechanical things and the creativity and artistry of the craftsman.

There is still much to be said for the mechanical wristwatch as a useful object. Whether it is self winding or manual, it doesn’t require batteries or plugging into USB ports, and it might keep track of any number of useful things. It’s also routine to make new ones waterproof. While quartz has more accuracy, for most purposes (including orbital mechanics navigation), mechanical watches are accurate enough it doesn’t matter. If it does matter, you can buy a hybrid quartz/mechanical self winding springdrive.  There is also the aspect of durability: if you take good care of them and avoid mishaps, most well made watches will continue to be serviceable without a major overhaul for … centuries. People hand them down to their grandchildren.

I expect there to be mechanical wristwatches made for as long as some remnant of Western Civilization continues to exist, if only to sell luxury products to the Chinese.  It’s a fundamental art form; a physical embodiment of the spirit of Western Faustian civilization.

I do not expect goofy innovations like the present form of “smart watches” to be around for as long. Smart watches are bugman technology.  They tell time … and do all kinds of other crap you don’t need such as informing you when you have email/slack updates, saving you the towering inconvenience of reading them a half second later on your phone or laptop. When you dump $600 on one of these goofy things, you can’t even expect it to be around in 20 years to give to the kids you (as a bugman) will never have, let alone 100 or 200 years as a $600 watch might. It isn’t because new “smart watches” have amazing new features which obsolete the old ones: it’s because the connectors and case will physically wear out and the operating system for your phone won’t support old models.

The difference between mechanical watches and smart watches is a sort of useful test case to generalize this sort of value judgement from. Consumerist capitalism has committed many great sins.   I could put up with most of them if they could get engineering aesthetics right. The world we live in is ugly.  Bugman engineering is one of the forms of ugliness which makes life more unpleasant than it needs to be.

Bugman devices are festooned with unnecessary LED lights. Whether it is a smoke alarm, a computer monitor switch, keyboard, power strip, DVD player, radio: you virtually never need an LED light to tell you that some object is hooked up to power. Especially objects which stay on all the time, like a smoke alarm or monitor. If you must have an indicator of activity; place a mechanical button on the object that makes a noise when you press it with power is on. Nobody is going to notice one among the sea of stupid little lights in a room have gone out. The time when it was “futuristic” to have little led’s all over your refrigerator or toaster is long past. Just stop it.

Bugman designed appliances have digital clocks you must set. There is no reason for your oven, blender, microwave, refrigerator, dish washer or water dispenser to know what time it is.  Power does go out on occasion (all the time in “futuristic” shit holes like Berkeley), and nobody wants to tell their stove what time it is. If you must have a clock; make one with a mechanical clock with hands you can easily move rather than navigating a 3 layer menu of membrane switches to set digits.

Bugman devices don’t use mechanical switches; they’re not “futuristic” enough. Capacitative switches are terrible and never work right. Touch screens on your car’s entertainment system are a horror. Membrane switches on your appliance or anything else are a planned obsolescence insult unless you are operating in a flammable or underwater atmosphere; the only reason to use membrane switches.

Bugman devices are besmirched with extraneous software and are networked when they don’t have to be. Being able to control your light bulb over wifi or bluetooth is almost never necessary. It is wasteful, a security nightmare and aesthetically disgusting. And no I don’t want my stove to be on the internet so its clock knows what time it is.

Bugman devices and services use invasive phone applications for payment instead of credit cards. If your device is hooked up to the internet enough to talk to a cell phone, it’s hooked up to the internet enough to use a credit card, crypto currency or paypal. Bugmen don’t mind the security and privacy nightmare of loading new executables on their nerd dildo phones.

Bugman devices complexify life and make people work rather than making their lives better. Every password, clock, networked device, app you have to manage, every battery you have to charge, change or replace is making your life worse. Bugman don’t care though; it helps fill the emptiness.

Bugman software substitutes software for actual experiences. Not all video games or online entertainment are bugman, but most VR applications or immersive social games (looking at you, Guitar Hero) are. Bugman sexuality; well, I bet they’re excited about sex robots.

Juicero, an internet equipped, phone interfacing, centrally planned/distributed subscription juice machine that costs $700 instead of a manual juicer that costs $10 and lasts multiple lifetimes.

Peloton: an internet equipped, exercise bicycle that costs $2000 plus subscription, as opposed to a $500 bike and some competitive friends.

Soylent is bugman food. It even looks like something the actual bug-man in the classic “The Fly” movie would eat. Hell, the bugmen in the media are trying us to get to eat actual bugs.

Many images and ideas from the excellent (arguably NSFW) “Shitty Future” twitter feed.

Ten predictions for 2030

Posted in Progress by Scott Locklin on January 27, 2020

Here’s a couple of predictions I’ll make for what happens in the next 10 years in science and technology.

 

  1. Computing and tech: Moore’s law (and Kryder’s law) will continue to fail. Computers in 2012 are about the same as computers in 2020 as far as I can tell; I’m typing this on a vintage 2011 laptop which is still actually faster than most 2020 laptops. I keep looking for server hardware that significantly (10x) beats a 2012 era Xeon box; please let me know if you can find any. Compiler engineers will become rare and in demand. The proliferation of programming languages will continue, with no clear winners. Operating systems will look much the same as they do now, with the possible exception of ideas from Mainframe days influencing modern time-sharing “clouds.” No major changes to computer UI; no virtual reality, no mass adoption of “wearables.”  ZNP systems become key pieces of technological infrastructure. Socially, the internet will continue to grow as an educational alternative to increasingly useless Universities. First mover effect will keep most existing tech companies in their present positions.
  2. Autonomous vehicles: There will be no stage-4, let alone stage-5 autonomous vehicles in 2030. “Driver” will still be the most common US job description. The whole thing always was a preposterous LARP; obviously so from the get-go; cruise control with collision avoidance and GPS is not autonomous driving. Rideshare companies will become federated taxicabs or fail.  The “gig economy” of underpaid servants will continue to substitute for manufacturing jobs for non college graduates in America.
  3. The “AI” apocalypse: There will be zero profitable companies based entirely around Deep Learning style machine learning. By 2030 there may be one profitable and large company besides Fair Isaac which is based on classic machine learning techniques such as boosting. Some older area of machine learning such as PGMs will come back into fashion as frameworks for GPUs are developed. No major disruptions to blue or white collar jobs.
  4. Data science: Significant breakthroughs in classical statistics will happen. Ideas such as sequential aggregation of experts or topological data analysis may form the nucleus of these new ideas. The new ideas will generally be ignored by people who aren’t in hedge funds. Arguably this is ongoing, and it is one of the few things I’m excited about for the future.
  5. Privacy: a high profile murder or kidnapping will result from the use of commodified adtech tracking as a service. It’s pretty trivial to track individuals based on their tracking bullshit now. The actual value of this for advertising is much lower than it is for intelligence work for most situations. While lawsuits are the main thing protecting people from this in the West, other civilizations have fewer scruples, and the internet is a global system. One with various only marginally regulated payment technologies. Cheap Chinese cameras and other “IoT” garbage, hacked or backdoored, will play a role in a significant political scandal.
  6. Quantum computing: The actual use case of Quantum Computing, aka factoring primes from very large integers, will not be achieved. The largest number legitimately factored by a quantum computer will be on the order of 1000. Not 1000 bits; something on the order of the number 1000; aka 10-11 bits. By “legitimately” I mean including all the gates needed to perform the Shor algorithm; about 72 * n^3. That’s on the order of 100,000 gates. RSA will continue to be used. “Quantum information theorists” will join sociology majors as a common form of training for underemployed baristas.
  7. Transportation/Energy Infrastructure: no significant changes. Electric cars will not become more than 5% of the US fleet. Electric aircraft efforts will fail.  Large scale passenger train efforts in the US will continue to fail due to bureaucratic sclerosis. No flying cars. No reusable spacecraft (single digit percent likelihood something like skylon works and is funded). “Renewables” will continue to grow in popularity without appreciably changing the US (or world) energy infrastructure. Some environmental disaster may be attributable to a renewable energy technology. No increased adoption of fission in the West. Imbeciles in 2030 will predict fusion breakthroughs by 2050.
  8. Engineering/manufacturing/mining: no major changes to manufacturing technology. Solid printing will continue to be a fringe and prototyping technology rather than a driver of new manufacturing techniques. Nanotech will continue in its nonexistence outside of bad science fiction plots. Fracking will grow in importance. Undersea mining will either create an ecological disaster or (hopefully) be banned before it does.
  9. Agriculture: agricultural productivity will continue to grow. Obesity and diabetes will become a worldwide problem.
  10. Scientific breakthroughs: none in physics, chemistry, astronomy, earth sciences. Some potential in biology (bioinformatics), medicine, paleontology and archaeology (via genetic and imaging technologies).

One of my firm beliefs as a technologist is the slow down in progress in the last 50 years. Had microelectronics not happened along, this would be abundantly obvious; most of the external technological world looks pretty much like it did 50 years ago. Now we have the situation where even microelectronics are not obviously getting much better, and this is a fact roundly ignored by most people who should know better. The disasters of coming years will mostly be from a lack of technological breakthroughs, and a lack of political recognition of this fact by the managerial classes who are living in a fantasy world where whiggism is as rational an ideology as it was 100 years ago. I assume this will result in continued “ghost dance” type crazes among managerial class types in the West. The present social unrest being experienced across the west is a result of this gross slow down in technological progress.  “The establishment” which has been in place across the West since 1945 are a technocratic elite. Their role was to increase productivity and distribute the fruits of this to the citizens of their countries. No productivity increase; no fruits to distribute; the social unrest will continue until new forms of social organization are found.

 

Speaking of which, I’m not terribly optimistic about new forms of social organization. While there are a few bright areas in terms of decentralized systems, by and large, the increase in surveillance technology and nation wide witch hunts via “social media” do not bode well. Countries are converging towards a sort of “social credit” system, which is plain old classism, abetted by surveillance technologies and enforced in the West via witch hunts and private enterprise. It’s a fragile and foolish system, to say nothing of being stupid and evil.

I could be wrong; we could spin up a sort of Manhattan project towards some achievable goal. A real one; not the Potemkin bureaucratic spoils nonsense that passes for such efforts in current year. There could be a large war which causes increased investment in technology. Doesn’t seem likely though; US can coast on its supremacy in military technology, the financial system and control of the seas for another 20 years at least, and the turds in charge will continue to think it some virtue of theirs, rather than the virtues of their grandparents.

Links to other people’s predictions:

https://futurebase.com/timeline/2030

https://www.quantumrun.com/future-timeline/2030

https://www.weforum.org/agenda/2016/11/8-predictions-for-the-world-in-2030/

https://www.news.com.au/lifestyle/real-life/wtf/time-traveller-man-from-2028-makes-specific-predictions-for-the-future/news-story/c410773ecf14d3b34df9deeeb5153b7a

A couple of dumb sugar plumb fairy predictions about 2020 (salted with a few good ones):

https://www.discovermagazine.com/mind/what-youll-need-to-know-in-2020-that-you-dont-know-now

https://www.wired.com/2000/07/the-year-2020-explained/

https://blogs.chicagotribune.com/news_columnists_ezorn/2010/01/2020.html

https://paulbuchheit.blogspot.com/2010/01/10-predictions-for-world-of-january-1.html

https://www.businessinsider.com/stratfor-predictions-for-the-next-decade-2010-1/china-doomed-1

 

The Fifth Generation Computing project

Posted in non-standard computer architectures by Scott Locklin on July 25, 2019

This article by Dominic Connor  reminds me of that marvelous artifact of the second AI winter (the one everyone talks about), “Fifth generation computing.”  I was not fully sentient at the time, though I was alive, and remember reading about “Fifth generation computing” in the popular science magazines.

It was 1982. Let’s not rely on vague recollections of what happened this year; some familiar things which happened that year. Tylenol scare, mandated breakup of the Bell System  and the beginning of the hard fade of Bell Labs, Falklands war, crazy blizzards, first artificial heart installed,  Sun Microsystems founded, Commodore 64 released. People were starting to talk about personal robotics; Nolan Bushnell (Atari founder) started a personal robot company. The IBM PC was released the previous year, somewhere mid year they had sold 200,000 of them, and MS-DOS 1.1 had been released. The Intel 80286 came out earlier in the year and was one of the first microprocessors with protected memory and hardware support for multitasking. The Thinking Machines company, attempting to do a novel form of massively parallel computing (probably indirectly in response to the 5thgen “threat”), would be founded in 1983.

Contemporary technology

The “AI” revolution was well underway at the time; expert system shells were actually deployed and used by businesses; Xcon, Symbolics, the Lisp Machine guys were exciting startups. Cyc -a sort of ultimate expert systems shell, would be founded a few years later. The hype train for this stuff was even more  lurid than it is now; you can go back and look at old computer and finance magazines for some of the flavor of it. If you want to read about the actual tech they were harping as bringing MUH SWINGULARITY, go read Norvig’s PAIP book. It was basically that stuff, and things that look like Mathematica. Wolfram is really the only 80s “AI” company that survived, mostly by copying 70s era “AI” symbolic algebra systems and re-implementing a big part of Lisp in “modern” C++. 

Japan was dominating the industrial might of the United States at the time in a way completely unprecedented in American history. People were terrified; we beat those little guys in WW-2 (a mere 37 years earlier) and now they were kicking our ass at automotive technology and consumer electronics. The Japanese, triumphant, wanted to own the next computer revolution, which was still a solidly American achievement in 1982. They took all the hyped technology of the time; AI, massive parallelism, databases, improved lithography, prolog like languages, and hoped by throwing it all together and tossing lots of those manufacturing-acquired dollars at the problem, they’d get the very first sentient machine. 

1) The fifth generation computers will use super large scale integrated chips (possibly in a non Von-Neumann architecture).
2) They will have artificial intelligence.
3) They will be able to recognize image and graphs.
4) Fifth generation computer aims to be able to solve highly complex problem including decision making, logical reasoning.
5) They will be able to use more than one CPU for faster processing speed.
6) Fifth generation computers are intended to work with natural language.

Effectively the ambition of Fifth generation computers was to build the computers featured in Star Trek; ones that were semi-sentient, and that you could talk to in a fairly conversational way.

 

People were terrified. While I wasn’t even a teenager yet, I remember some of this terror. The end of the free market! We’d all be Japanese slaves! The end of industrial society! DARPA dumped a billion 1980s dollars into a project called the Strategic Computing Initiative in an attempt to counter this (amusingly one of the focuses was … autonomous vehicles -things which are still obviously over the rainbow). Most of the US semiconductor industry and main frame vendors began an expensive collaboration to beat those sinister Japanese and prevent an AI Pearl Harbor. It was called the Microelectronics and Computer Technology Corporation (MCC for some reason), and it’s definitely ripe for some history of technology grad student to write a dissertation on it beyond the Wikipedia entry.  The Japanese 5th gen juggernaut was such a big deal, the British (who were still tech players back then) had their own copy of this nonsense, called the “Alvey Programme” -they dumped about a billion pounds in todays money into it. And not to be left out, the proto-EU also had their own version of this called ESPIRIT with similar investment levels. 

 

Prolog was of course the programming language of this future technology. Prolog was sort of the deep learning of its day; using constraint programming, databases (Prolog is still a somewhat interesting if over -lexible database query language), parallel constructs and expert system shell type technology, Prolog was supposed to achieve sentience. That’s not worked out real well for Prolog over the years: because of the nature of the language it is highly non-deterministic, and it’s fairly easy to pose NP-hard problems to Prolog. Of course in such cases, no matter how great the parallel model is, it still isn’t going to answer your questions.

 

One of the hilarious things about 5th generation computers is how certain people were about all this. The basic approach seemed completely unquestioned. They really thought all you had to do to build the future was take the latest fashionable ideas, stir them together, and presto, you have brain in a can AI. There was no self respecting computer scientist who would stand up and say “hey, maybe massive parallelism doesn’t map well onto constraint solvers, and perhaps some of these ambitions, we have no idea how to solve.” [1] This is one of the first times I can think of an allegedly rigorous academic discipline collectively acting like overt whores, salivating at the prospects of a few bucks to support their “research.” Heck, that’s insulting to actual whores, who at least provide a service.

 

 

Of course, pretty much nothing in 5th generation computing turned out to be important, useful, or even sane. Well, I suppose VLSI technology was all right, but it was going to be used anyway, and DBMS continue to be of some utility, but the rest of it was preposterous, ridiculous wankery and horse-puckey. For example; somehow they thought optical databases would allow for image search. It’s not clear what they had in mind here, if anything; really it sounds like bureaucrats making shit up about a technology they didn’t understand. For more examples:

“The objective stated (Moto-oka 1982 p.49) is the development of architectures with particular attention to the memory hierarchy to handle set operations using relational algebra as a basis for database systems. “
“The objective stated (Moto-oka 1982 p.53) is the development of a distributed function architecture giving high efficiency, high reliability, simple construction, ease of use, and adaptable to future technologies and different system levels.”
“The targets are: experimental, relational database machine with a capacity of 100 GB and 1,000 transactions a second; practical, 1,000 GB and 10,000 transactions a second. The implementation relational database machine using dataflow techniques is covered in section 8.3.3.”
“The objective stated (Moto-oka 1982 p.57) is the development of a system to input and output characters, speech, pictures and images and interact intelligently with the user. The character input/output targets are: interim, 3,000-4,000 Chinese characters in four to five typefaces; final, speech input of characters, and translation between kana and kanji characters. The picture input/output targets are: interim, input tablet 5,000 by 5,000 to 10,000 by 10,000 resolution elements; final, intelligent processing of graphic input. The speech input/output targets are: interim, identify 500-1,000 words; final, intelligent processing of speech input. It is also intended to integrate these facilities into multi-modal personal computer terminals.”
“The Fifth Generation plan is difficult and will require much innovation; but of what sort? In truth, it is more engineering than science (Fiegenbaum & McCorduck 1983 p 124). Though solutions to the technological problems posed by the plan may be hard to achieve, paths to possible solutions abound.” (where have I heard this before? -SL)

The old books are filled with gorp like this. None of it really means anything.  It’s just ridiculous wish fulfillment and word salad.  Like this dumb-ass diagram:

 

There are probably lessons to be learned here. 5thGen was exclusively a top down approach. I have no idea who the Japanese guys are who proposed this mess; it’s possible they were respectable scientists of their day. They deserve their subsequent obscurity; perhaps they fell on their swords. Or perhaps they moved to the US to found some academic cult; the US is always in the market for technological wowzers who never produce anything. Such people only seem to thrive in the Anglosphere, catering to the national religious delusion of whiggery.

Japan isn’t to be blamed for attempting this: most of their big successes up to that point were top-down industrial policies designed to help the Zaibatsus achieve national goals. The problem here was … no Japanese computer Zaibatsu worth two shits which had the proverbial skin in the game -it was all upside for the clowns who came up with this, no downside.  Much like the concepts of nanotech 10 years ago, quantum computing or autonomous automobiles now; it is a “Nasruddin’s Donkey bet” (aka scroll to bottom here) without the 10 year death penalty for failure.

Japan was effectively taken for a ride by mountebanks. So was the rest of the world. The only people who benefited from it were quasi-academic computer scientist types who got paid to do wanking they found interesting at the time.  Sound familiar to anyone? Generally speaking, top down approaches on ridiculously ambitious projects, where overlords of dubious competence and motivation dictate a  R&D direction that don’t work so well; particularly where there is software involved. It only works if you’re trying to solve a problem that you can decompose into specific tasks with milestones, like the moon shot or the Manhattan project, both of which had comparatively fairly low risk paths to success. Saying you’re going to build an intelligent talking computer in 1982 or 2019 is much like saying you’re going to fly to the moon or build a web browser in 1492. There is no path from that present to the desired outcome. Actual “AI,”  from present perspectives, just as it was in 1982, is basically magic nobody knows how to achieve. 

Another takeaway was many of the actual problems they wanted to solve were done in a more incremental way while generating profits. One of the reasons they were trying to do this was to onboard many more people than had used computers before. The idea was instead of hiring mathematically literate programmers to build models, if you could have smart enough machines to talk to people and read charts and things the ordinary end user might bring to the computer with questions, more people could use computers, amplifying productivity. Cheap networked workstations with GUIs turned out to solve that in a much simpler way; you make a GUI, give the non spergs some training, then ordinary dumbasses can harness some of the power of the computer. This still requires mentats to write GUI interfaces for the dumbasses (at least before our glorious present of shitty electron front ends for everything), but that sort of “bottom up, small expenditures, train the human” idea has been generating trillions in value since then.

The shrew-like networked GUI equipped microcomputers of Apple were released as products only two years after this central planning dinosaur was postulated. Eventually, decades later, someone built a mechanical golem made of microcomputers which achieves a lot of the goals of fifth generation computing, with independent GUI front ends. I’m sure the Japanese researchers of the time would have been shocked to know it came from ordinary commodity microcomputers running C++ and using sorts and hash tables rather than non-Von-Neumann Prolog supercomputers. That’s how most progress in engineering happens though: incrementally[2]. Leave the moon shots to actual scientists (as opposed to “computer scientists”) who know what they’re talking about. 

 

1988 article on an underwhelming visit to Japan.

1992 article on the failure of this program in the NYT.

 

[1] Some years later, 5 honest men discussed the AI winter upon them; yet the projects inexorably rolled forward. This is an amazing historical document; at some point scholars will find such a thing in our present day -maybe the conversation has already happened. https://www.aaai.org/ojs/index.php/aimagazine/article/view/494 … or PDF link here.

[2] Timely Nick Szabo piece on technological frontiersmanship: https://unenumerated.blogspot.com/2006/10/how-to-succeed-or-fail-on-frontier.html