12 September 2013

My opinion, literally

Lexicography

Previously

I wrote this a few weeks back, in response to something my friend Michael wrote, but I was lazy about posting it.

It's about the Merriam-Webster dictionary defining "literally" to mean "figuratively". See slate.

My opinion

The dictionary has to be descriptive, not prescriptive. It should reflect how people have actually spoken and used words. Ultimately, that's what language is.

But by the same token, people actually use dictionaries prescriptively. They turn to dictionaries for authority on the "right" meaning of words.

I would take a middle position. A dictionary must eventually track usage, but there's no need for it to rush to anoint every popular solecism.

What lexicographers do it collect a corpus of contemporary usage and then group the words according to word sense, as they see it. I'm not surprised that they found so many hyperbolic "literally"s. I'm sure they also had access to literally tons of people who felt figurative "literally" to be a solecism.

There's merit, as some lexicographers do, in characterizing these groups of words in more sophisticated ways. Hyperbolic senses can be noted. So can loose senses ("sarcasm" that lacks irony) and apparent substitutions ("effect" where "affect" is meant).

It's too bad Merriam-Webster stopped before doing that, and I think they deserve all the criticism for it.

Patent Fix 1

Patent Fix 1

Previously

Jan Wolfe blogs Patent defendants aren't copycats. So who's the real inventor here?.

Robin Hanson also writes about this. While his central illustration is somewhat implausible, it nevertheless puts the issue on a concrete footing. Briefly, a hypothetical business finds better routes for city drivers for a price. Then they want to forbid all the drivers in the city from driving "their" route without a license. (This example is apparently set in a world where nobody every heard of Google maps, Mapquest, or even paper maps.)

Also see Defending reinvention, So How Do We Fix The Patent System?, and A Call for an Independent Inventor Defense.

Recent changes in patent law don't seem to have addressed reinvention.

Some points of consensus

We all seem to agree that the biggest issue is patent protection vs re-invention. If re-invention was somehow metaphysically impossible, the patent situation would be more defensible1.

We also seem to agree that software patents aggravate this problem.

This is a problem that I have been kicking around for a while too. I have, I think, a different analysis to offer and something of a solution.

The deep problem

The deep problem, as I see it, is a problem is common to all intellectual pseudoproperty. First the discoverer sells it - to an individual or to society via the social contract of patenting. Only afterwards can the invention be seen by the buyer and evaluated.

For the individual, such as the driver licensing a better route in Robin's example, this makes the route an experience good - he can only tell whether it's worth what he paid after he receives it. Then he can't return the good by unlearning it.

He may find that it's not worth what he paid because it's bad. More relevant to patent re-invention, he may find that he already knew of that route. Perhaps he sometimes uses it but mostly prefers a slightly longer scenic route. He shouldn't lose use of a route he already knew just because he "bought" the same from this business and then returned it.

This at least could be solved by contract - perhaps the driver can get a refund if he proves that he already knew that route. For society, it's worse. Patent law basically throws up its hands at the question of what the loss of opportunity for re-invention costs. It deems the cost to be zero, which is simply insane.

Why it's tricky to solve

It's sometimes said that the invention of the eraser on the pencil was obvious - after it was invented; before that, nobody could think of that. As it turns out, that's questionable2 for the pencil-eraser but the general point stands.

So we don't want re-invention to consist of stating, after one has seen the invention, "Why, it's obvious to anybody". That's hindsight bias. That's cheating. We want to measure problem hardness in foresight. How hard did it appear before you knew the answer?

So how can we measure problem hardness?

For any patent, there is some problem that it solves. This isn't just philosophical; a patent application must contain this. It's part of one of the less formal sections of an application, but even so, it's made explicit for every patent the USPTO ever granted.

Imagine a public blackboard where anybody can write down a problem. Our would-be inventor writes down the problem that his as-yet undisclosed invention solves. He leaves the problem on the board for a year. After a year, nobody has written down the solution.

Our inventor then says "Given an entire year, nobody has solved this problem except me. Therefore even if my invention is granted patent protection, nobody has lost the chance to invent it themselves. Since opportunity for re-invention was the only thing anybody really stood to lose, I should be granted patent protection. I stand ready to reveal my solution, which in fact solves the problem. If I'm granted patent protection, everybody wins."

We're not assuming that our inventor has already invented the invention. He could pose a problem that he feels pretty sure he can solve and only start working on it later in the year.

There are of course holes in that solution. Let's try to fill them.

Issues

A dialog about attention

"Not so fast". (Now we need names. I will call the first inventor Alice and the new voice Bob)

"Not so fast" says Bob. "I could have solved that problem, but I was working on something else more important. I'm confident that when the day comes that I actually need a solution, I can solve it myself. I like that better than paying your license fee. So your patent protection does deprive me of the opportunity for re-invention."

"You haven't proved you can do that." replies Alice. "Nobody answered my challenge, so why should I believe you?"

"Correction: they didn't answer it for free." says Bob. "It does take non-zero time and attention - and by the way, so would using your patent even if the license fee was $0"

Bob continues "Maybe other possible inventors felt the same as I did. Or maybe they're just not paying attention to that blackboard. If everybody has to pay attention to everything you write on that blackboard, that imposes a public cost too."

"I'll tell you what: Offer a reasonable bounty to make it worth my time, say $5000, and I will show you my solution to your problem."

"I haven't got $5000 to spend on that" says Alice, "I'm a struggling independent inventor with one killer idea in my pocket. And if I did have $5000 then I'd just be paying you to reveal your invention. I already have one, so the value to me is $0. If you're not bluffing, I'll be down $5000."

"If you don't have to offer a reward, I can play that game too," replied Bob, "but I won't leave it at one problem. I'll write write down thousands of them. Then I'll own all the problems that nobody paid attention to, many of which will be actually quite easy, just ignored. I'll solve some of them and nobody can use the solutions without licensing them from me."

"I see where that's going" says Alice. "I'd do the same." A look of absolute horror crossed her face. "We'd be back in the Dark Ages of the US patent system!"

How do we direct attention reasonably?

Collectively, Alice and Bob have a pretty good idea what's easy to solve and what's not. The trick is truthfully revealing that knowledge at a practical cost.

One possible solution quickly suggests itself to those acquainted with idea markets: We can adapt Hanson's old lottery trick.

I'll start by explaining the lottery. We'll go beyond that later. The idea is that Alice and everybody else who writes a problem on the blackboard pays a small license fee. The whole amount - let's say it's still $5000 - is offered on one problem chosen by lottery. Maybe it's offered on a few of them, but in any case a small fraction of what there is.

That's great for that one problem, but it leaves the other problems unmeasured.

That's where decision markets come in. Investors place bets on whether a problem, should it be chosen for a bounty, will in fact be solved. Then one particular problem is chosen by lottery. The bets about its hardness are settled while the other bets are called off. It's easy for the system to see whether someone has claimed the bounty. We won't tackle the quality of the solution until later in this post.

The hardness-market price determines whether the problem is considered easy or not - there may be finer gradations but we won't get into that here.

So we've amplified the signal. By offering a bounty on one problem, we've measured the hardness of all the problems. We'll improve on this later.

More dialog

So this scheme is implemented. A few weeks later, Bob comes storming in.

"Somebody wrote on the blackboard the exact problem that I've been working on!"

"That's odd." says Alice

"Did you do this to get back at me, Alice?"

"How does that inconvenience you? You could earn $5000."

"My invention is worth millions! Now I have to disclose it for a one-time payment of $5000? That doesn't even cover my research and development costs!"

"Well, you can hardly expect anyone to offer a million dollars for a bounty."

"That's true. Still, this is very wrong. Since you wrote the problem, you should be required to have a solution too."

"It just so happens that I do, but if I disclosed it, you'd just copy it. You should have to show your solution first."

"Then you'd copy it. See, Alice, I don't trust you either."

"I really had no idea you were working on it too, Bob. But if you really do have a million-dollar invention too, why should either of us sell it for $5000? As far as we know, only the two of us have it. Why should the two of us collectively get less than it's worth?"

"Luckily we found out in time. That's just sheer luck. We could agree to ignore the bounty and split the rewards of the patent between us."

"50/50"

The two inventors shook hands and then compared notes.

"Hey!" exclaimed Bob "I shouldn't have assumed that just because your invention solved the same problem, it was as good as mine. It's not! Mine's cheaper to manufacture. I'd have got about 95% of the market share."

"No, I'd have beat you. Mine's more stylish and easier to use."

The two inventors glared at each other, each convinced they had gotten the worst of the deal.

Bugs

So we have the following bugs:

  • The bounty acted like a price, but wasn't really a sensible price. In fact, it didn't even try, it just set one fixed price for everything.
  • The bounty overrode the market payoffs, which are a better measure of invention quality.
  • Relating the hardness to the number of times the bounty is claimed measures the wrong thing. What we need is solution unobviousness. This is trickier, since we have to look past the expected number of solutions and see how many actual solutions overlap.
  • If the first inventor to disclose gets a patent monopoly, it's unfair to Alice, who posed the problem and paid to do so. It shuts her out of her own invention.
  • If the first to disclose doesn't get a patent monopoly, for inventions whose expected market value is more than the bounty, the likelihood of the bounty being accepted will be too low. We'll see fewer than we should and therefore we'll underestimate obviousness.

Second draft

In order to fix those bugs, we're going to redo the bounty system:

  • Claiming the bounty doesn't require openly disclosing an invention. It requires only an unforgeable commit, probably as a secure hash of the disclosure document. The reveal and the payment will come later. NB this allows the original problem-poser to participate in the race for the solution.
  • After a predetermined period of time:
    • The claimed inventions are to be disclosed (all at once)
    • Inventions not disclosed within a given time-frame are ineligible.
    • Disclosures that don't match the prior commit are ineligible.
    • Unworkable inventions are ineligible.
    • Each distinct workable invention gets patent protection
      • Technically, that's each distinct patent claim - a USPTO patent application typically contains many related patent claims.
      • If multiple inventors invented the same invention, they each get an equal share. This is suboptimal vs sockpuppets, though.
      • Here I'm assuming that novelty against prior art isn't an issue. It won't be much of one, because why would Alice pay money to pose a problem whose solution is already publicly known? We can just say that the existence of a prior invention voids a patent claim, just like now.
    • Each distinct workable invention gets an equal fraction of the bounty.
  • Non-bountied inventions are treated essentially the same way, just minus the bounty and the bounty time period.
  • The market pays off in proportion to the count of unique workable solutions, rather than the count of all solutions.
    • We don't want to use some sort of "density of uniqueness", because that's too easily spoofed by submitting the same solution many times.
  • To estimate solution unobviousness:
    • For bountied problems, we use a count of solutions, backed off somewhat towards the market's prior estimate.
    • For non-bountied problems, we directly use the market's prior estimate.

Measuring invention workability and quality

We still need a way to measure invention workability. This is traditionally the job of a patent examiner. However, we've piled more work onto them, and there are concerns about how good a job they do. This post is already long, though, so I won't try to design a better mechanism for that here.

Progressively increasing bounties

One way to aim bounties more efficiently is to start by offering small bounties. Then for some of those problems whose bounties were not claimed, raise the bounty. That way we are not overpaying for easy answers.

Time frame

We haven't addressed the question of how long the problem period should be. We may want it to work differently when there is a bounty, since then we need a standard measuring period.

It's not hard to propose. I could say "six months", but I want to leave it open to more flexible rules, so I won't propose anything here.

Relates to how long a patent monopoly should be in force

This mechanism also provides a somewhat more principled way of deciding how long a patent should be in force: it should relate to how long the problem period is. Perhaps the two should be linearly proportional.

Bug: We charge for posing a problem

Posing a problem well is valuable and is a task in and of itself. Yet we've charged the problem-poser for the privelege. This isn't good, and I'd like it to be the other way.

We could try to recurse, with the problem being: "What are some unsolved problems in field X?" but then the solution is no longer in a standard form as formal patent applications are.

This post is already long, so I will leave it at that.

Footnotes:

1 I will mention in passing that AIUI re-invention is allowed but only under stringent conditions that are only practical for well-heeled institutions. The slightest exposure to a patent "taints" a would-be rediscoverer forever. IANAL so take this with a grain of salt.

2 That patent was later disallowed because it was a mere combination of two things, which is not patentable. See Eraser. Regardless, the general point stands.