- Paul Christiano has bet me $500 at even odds that a self-driving car can be reliably hailed by a member of the general public in at least 10 North American cities by July 2023.
Details: At least 8 cities must be outside San Fransisco Bay Area. The car must available on at least 50% of days, i.e., not confined to very narrow weather or traffic situations. The car must be self-delivering, in the sense that it drives itself to the user, but not necessarily fully self-driving, in the sense that the user might need to drive it to the destination. (It’s easy to imagine tech and regulatory scenarios where self-driven cars are limited to speeds that are unacceptably slow during transportation of passengers, like the ~20 mph that Google’s car usually does, but are sufficient for getting to the hailing passenger if the density is high enough.) Carl Shulman will adjudicate any edge cases.
I ascribe a 45% chance that a self-delivering car reaches this threshold, and 38% chance that a fully self-driving car does.
Here’s a list of optimistic predictions for self-driving car timelines, which notably doesn’t mention the recent Google pessimism.
- People I know build great stuff!
My brother Will is an electrical engineer at Apple. He has been heavily involved in improving Apple display technology for the past two years, especially the True Tone feature and especially with the iPad Pro 9. Well, the reviews from the experts are in:
The Absolute Color Accuracy of the iPad Pro 9.7 is Truly Impressive as shown in these Figures. It is the most color accurate display that we have ever measured. It is visually indistinguishable from perfect, and is very likely considerably better than any mobile display, monitor, TV or UHD TV that you have.
Unrelatedly, Alex Flint is a great friend of mine who’s startup Kite just launched a private beta. The response at HackerNews for their coding copilot was damn enthusiastic.
Congratulations!
- Aircraft Carrier USS Carl Vinson installs UAV operating center for commanding drone tankers. The drone tankers will fill a refueling role currently carried out by the manned F/A-18E and F/A-18F aircraft. I didn’t even know the Navy had been landing drone combat aircraft on their carriers for several years:
The videography isn’t that good, but check out a wider angle at 3:30. - George Church: “It is arguably the most clever CRISPR gadget to date.”
Harvard University biochemist David Liu and post-doctoral fellow Alexis Komor, who led the work, have filed a provisional patent application on their invention, which changes one-letter misspellings in DNA called point mutations….“Most known human genetic variations associated with disease are point mutations,” said Liu. “Current gene-editing methods aren’t particularly good at correcting those.”…The problem addressed by the new technique is not the usual one discussed with CRISPR, namely, off-target effects. That refers to changing a region of the genome other than the intended one. While early genome-editing experiments had that problem, there has been “tremendous progress” in fixing it, said Dr. Keith Joung of Massachusetts General Hospital…The workhorse of the CRISPR system is an enzyme that cuts DNA. In some cases, simply cutting out a disease-causing gene might be enough to achieve a cure, as with cancer-causing genes. If deleting a gene isn’t sufficient, and a replacement gene is also required to treat a particular disease, CRISPR can carry substitute DNA, such as a healthy cystic fibrosis gene….But genomes don’t take kindly to being cut. When the CRISPR enzyme whacks the double helix, the cell tries to “get the broken ends back together,” Liu said. Molecules in the cell grab the four basic components of DNA — represented by the letters A, T, C, and G — from the cellular soup and cram them into the cleaved DNA like someone pushing spackle into the crack of a Ming vase; other molecules cut out segments of DNA….For nearly two years he, Komor, and their colleagues tried to improve that aspect of CRISPR. They replaced its usual cutting enzyme with a dud called “dead Cas9.” Like blunt scissors biting down on fabric, dead Cas9 can latch on to DNA but not cut it. Then they attached two other proteins that change one DNA letter to another and lock it in place. So far, of the 12 possible changes (A to T, C to G, G to C …), the new system can make two: C to T and G to A.
(Journal article.) Incidentally, here’s a
long interview by Edge with George Church on Crispr. (H/t Ryan Carey.) - The LHC starts taking data again on Monday. When it does, you can watch live ATLAS events.
- The best popular level article explaining the reproducibility crisis. (Not much new, but a good go-to link.) Physics correctly does not get off the hook:
And yet…deductive, theory-building physics in the mold of Newton and Lagrange, Maxwell and Einstein, is a tiny fraction of modern science as a whole. In fact, it also makes up a tiny fraction of modern physics. Far more common is the delicate and subtle art of scouring inconceivably vast volumes of noise with advanced software and mathematical tools in search of the faintest signal of some hypothesized but never before observed phenomenon, whether an astrophysical event or the decay of a subatomic particle. This sort of work is difficult and beautiful in its own way, but it is not at all self-evident in the manner of a falling apple or an elliptical planetary orbit, and it is very sensitive to the same sorts of accidental contamination, deliberate fraud, and unconscious bias as the medical and social-scientific studies we have discussed. Two of the most vaunted physics results of the past few years—the announced discovery of both cosmic inflation and gravitational waves at the BICEP2 experiment in Antarctica, and the supposed discovery of superluminal neutrinos at the Swiss-Italian border—have now been retracted, with far less fanfare than when they were first published.
This is clever research design I hadn’t heard:
In a survey of two thousand research psychologists conducted in 2011, over half of those surveyed admitted outright to selectively reporting those experiments which gave the result they were after. Then the investigators asked respondents anonymously to estimate how many of their fellow scientists had engaged in fraudulent behavior, and promised them that the more accurate their guesses, the larger a contribution would be made to the charity of their choice. Through several rounds of anonymous guessing, refined using the number of scientists who would admit their own fraud and other indirect measurements, the investigators concluded that around 10 percent of research psychologists have engaged in outright falsification of data, and more than half have engaged in less brazen but still fraudulent behavior such as reporting that a result was statistically significant when it was not, or deciding between two different data analysis techniques after looking at the results of each and choosing the more favorable.
The biggest flaw with this piece is when it laments that “vulnerable young” theories must weather “an initial storm of skepticism”, without making it clear that there are uncountable new young theories that turn out to be very wrong, and thus justly deserving of criticism. And that, furthermore, the amount of resources spent on bad new theories is significant, making this critical process important.
This bit was biting but on-the-money:
…science was unprepared for the blossoming of the Cult of Science. The Cult is related to the phenomenon described as “scientism”; both have a tendency to treat the body of scientific knowledge as a holy book or an a-religious revelation that offers simple and decisive resolutions to deep questions. But it adds to this a pinch of glib frivolity and a dash of unembarrassed ignorance. Its rhetorical tics include a forced enthusiasm (a search on Twitter for the hashtag “#sciencedancing” speaks volumes) and a penchant for profanity. Here in Silicon Valley, one can scarcely go a day without seeing a t-shirt reading “Science: It works, b—es!” The hero of the recent popular movie The Martian boasts that he will “science the sh— out of” a situation. One of the largest groups on Facebook is titled “I f—ing love Science!” (a name which, combined with the group’s penchant for posting scarcely any actual scientific material but a lot of pictures of natural phenomena, has prompted more than one actual scientist of my acquaintance to mutter under her breath, “What you truly love is pictures”).
- Relatedly, Scott Alexander: “Maybe the most important article I’ve read this year: “Statistically Controlling for Confounding Constructs Is Harder than You Think“. On the same page, Scott also points to this blog post by Gelman: “Too big to fail: Why it’s unrealistic to expect scientific journals to retract their huge backlog of erroneous papers“. And earlier in the month, The ideology is not the movement:
It continues to worry me that people who would never display an American flag on their lawn because “nations are just a club for hating foreigners” now have a campaign sign on their lawn, five bumper stickers on their car, and are identifying more and more strongly with political positions – ie clubs for hating their fellow citizens.
- I did not realize that prion diseases can be so infectious within animals. Saliva is enough. (Its difficult to transmit between humans without cannibalism or blood contact.)
- Fermat’s library highlights a paper every week for community commenting. Here’s Toth’s “New proof of a minimum property of the regular n-gon“. Pretty nice PDF-comment-overlay interface.
- List of examples of Stigler’s law: the supposed tendency of eponymous expressions for scientific discoveries to honor people other than their respective originators.
- Good outline by Holden of Open Phil’s high-risk high-reward philosophy. Notable section:
We don’t: avoid controversial positions or adversarial situations. All else equal, we would rather not end up in such situations, but making great effort to avoid them seems incompatible with a hits-based approach. We’re sympathetic to arguments of the form, “You should be less confident in your position when intelligent and well-meaning people take the opposite side” and “It’s unfortunate when two groups of people spend resources opposing each other, resulting in no net change, when they instead could have directed all of their resources to something they agree on, such as directly helping those in need.” We think these arguments give some reason to prefer GiveWell’s top charities. But we feel they need to be set aside when aiming for “hits.”
I am generally very skeptical of EAs drawn to overtly political causes, but this is a good paragraph that acknowledges many of the best arguments. Still, it means I can’t assume that I am roughly aligned with Open Phil’s values. They have made it clear in the past that they don’t put any intrinsic weight on freedom vs. coercion.
The real question will be whether they follow the implications of their justification. That is, will they avoid getting into deadlocked resource wars, and concentrate where the meta explanations for confidence in the face of disagreement are strongest? Alternatively, will they be drawn toward interventions popular with their political tribe?
Relatedly, GiveWell has announced what reads to me as a major shift in focus (although people following closely will probably say this transition has been underway for at least a year or two):
…we have come to believe that the kind of work we’ve recently been doing to find top charities – deeply investigating the most promising-seeming charities we know of, based largely on which interventions they carry out – has limited promise. In past years – and at the beginning of this year – we hoped that these investigations would lead relatively quickly to new top charities. Now, we believe that we’ve already (previously) identified most of the strongest charities by our criteria, and there aren’t many strong candidates left (though there are a few that we continue to investigate, and we remain willing and eager to investigate further promising groups if we come across them). With that in mind, we have begun seeing more potential in other research priorities, such as supporting the development of new organizations and new evidence bases.
- We plan to focus much of our capacity on a small number of initiatives that are unlikely to result in new top charities in 2016, but which we hope will lead to new top charities that are competitive with our current top charities in 2017 or 2018.
- We plan to intensify our work following our current top charities and are tentatively planning to make site visits to distributions funded by the Against Malaria Foundation and work supported by Evidence Action’s Deworm the World Initiative.
- We are also planning a substantial project focused on the question of whether or not we should recommend that Good Ventures give significantly more than it has in the past to support insecticide-treated nets, arguably the most promising area we know of for substantial additional funding.
- We also hope to take on additional work (described in detail below) but plan to prioritize this work below the items listed above.
- We plan to put more staff time into donor outreach than we have in the past and discuss our priorities for that work below.
- A comparison tool for meal-replacement drinks. There are a lot.
- SpaceX sticks the drone-ship landing, in 4K resolution…
…and from the rocket’s perspective:
Onboard view of landing in high winds pic.twitter.com/FedRzjYYyQ
— SpaceX (@SpaceX) April 9, 2016
The rocket that landed back on the ground three months will be kept permanently as a monument on the SpaceX campus, but this ship-landed rocket is expected to fly again in a few months for a paying customer. Additionally, Musk intends to announce at a conference in September the details on SpaceX’s plan for a Martian city. Both of those points made in this otherwise uneventful interview:
https://www.youtube.com/watch?v=VNygOavo2mY
Something that’s new to me:SpaceX indicated in November 2014 that they plan to eventually use ASDS as a floating launch platform, refueling a landed first stage with sufficient propellant to enable it to fly back to its launch site. No date has been provided for when this conceptual capability might be developed, tested, or made operational, or if it is still being actively developed.
- Lyft spends big on user acquisition and gets modest success. I am cautiously optimistic that we will reach an equilibrium of competition. The ability for riders and drivers to run multiple services simultaneously substantially reduces networks effects in comparison to things like Facebook. The biggest risk would be for Uber to win out nationwide, and then make an implicit but credible dumping threat — reducing prices below cost — in any city that a new company tried to enter.
- Reminder: Warren Buffet is still winning his 10-year long bet about index funds.
- Minecraft and modern play. Not much new, but enjoyable to read, especially if you like hartwarming stuff like this:
One day last fall, I visited Gus, a seventh grader in Brooklyn. He was online with friends on a server they share together, engaging in boisterous gladiatorial combat. I watched as he typed a command to endow himself with a better weapon: “/give AdventureNerd bow 1 0 {Unbreakable:1,ench:[{id:51,lvl:1}],display:{Name:“Destiny”}}.” What the command did was give a bow-and-arrow weapon to AdventureNerd, Gus’s avatar; make the bow unbreakable; endow it with magic; and name the weapon Destiny, displayed in a tag floating over the weapon. Gus had plastered virtual sticky-notes all over his Mac’s desktop listing the text commands he uses most often. Several commands can be packed into a “command block,” so that clicking on the block activates them, much as clicking on a piece of software launches it.
There is some pontificating too:
Schools and governments have spent millions on “let’s get kids coding” initiatives, yet it may well be that Minecraft’s impact will be greater.
May?
- Is this Caplan critique of Mill correct, localized as it may be?
- Fascinating historical Wikipedia article on Le Sage’s theory of gravitation. (H/t Godfrey Miller.)
- Good news roundup from MIRI, including a pointer to a new paper by Bostrom on AI openness.
- You know all those ridiculous renderings of skyscrapers covered in trees? One of them actually got built in Milan.
The real photos don’t quite live up to the CG, but it’s not bad. (HN discussion.)
- UN peacekeepers are disproportionally staffed by poor southeast Asia because the pay is good. (Article buries the lede.)
- HistPhil, a blog on the history of philanthropy, is partially funded by the Open Philanthropy Project.
- Chip escapes Japanese zoo, climbs electrical lines, is shot with dart gun, then falls into waiting rescue-blanket trampoline a few minutes later when the sedatives kick in. Everything but the first part caught on video.
- Steve Hsu covers Geoff Hinton’s talk on Deep Learning.
- Question prompted by Russian fighter planes buzzing a US warship below the height of its superstructure: is there a sensible game-theoretic justification for this behavior, or is this literal insanity? Should large corporations being doing this to maximize shareholder value? (Note: “The USS Donald Cook was stationed outside Kaliningrad, a Russian exclave sandwiched between NATO members Lithuania and Poland where Moscow deployed state-of-the art Iskander missiles last year….Its proximity was no coincidence, according to Bronk, who said it was ‘just as calculated by the Americans’ as the Russian jet flyby — albeit far less immediately dangerous.”)
- Popular coverage of Randall’s updated paper on double-disk dark matter. We’ll be citing this stuff as one scenario where decoherence detection would be very useful, but these sort of comments should throw up all sorts of alarm bells:
They were trying to explain an apparent Milky Way anomaly. Known as the “Fermi line,” it was an excess of gamma rays of a certain frequency coming from the galactic center. “Ordinary dark matter wouldn’t annihilate enough” to produce the Fermi line, Randall said, “so we thought, what if it was much denser?” The dark disk was reborn. The Fermi line vanished as more data accumulated, but the disk idea seemed worth exploring anyway.
Sunk costs!
- Reinhard Werner on the bad effects of chasing impact factors. Mostly well-known complaints, and I wish he had spent more time thinking about what might actually be changed in the real world rather than just listing the problems (which are legion). But I did like this
Some people in that field have thought that one could use the electronic supplements [in a PRL] to give the supporting arguments or proofs for claims made in the body of a letter. Indeed this would be a good paper format, and could make PRL attractive as a place to publish serious theory. But editorial policy seems to be against that, because letters are supposed to stand on their own, and referees typically don’t appreciate it. This means that PRL style “top quality” is largely denied to fields like argument-rich theoretical or mathematical physics, where the pudding is in the proof.
I have always thought that any paper over, say, 6 pages, should typically have large chunks relegated to appendices which are longer than the paper. I believe that the most logical format for a scientific result is in a hierarchy of increase length and detail: title, abstract, body, appendices. (Indeed, many people read papers like this even if they aren’t structured that way; first read the introduction, conclusion, and maybe a few figures, then decide whether to read the main body.) Anything that can be broken off into the appendix should be.
- The Sudbury Neutrino Observatory (SNO) contained 1,000 tones of heavy water, which is valued at $300 million. Los Alamos loaned the water to SNO for $1, but the insurance cost $1 million per year.
- A representation of everything that’s wrong with Academia, packaged as a new and exciting research technique?
Last week, the SFI postdocs embarked upon a scientific experiment in collaboration and deadline compression. They agreed to spend three days with each other, away from the institute, focusing on one scientific topic (not discussed in advance), which they could all contribute to from their various fields of study and would result in interesting and publishable results. The paper titled: Dynamics of beneficial epidemics was produced, from conception of idea, to execution, to writing, by a team in just 72 hours and was submitted at precisely 72 hours after the experiment began.
The resulting arXiv paper. Gizmodo article. (H/t Matteo Smerlak.)
- Good stuff from Tyler: The 6 elements of an effective apology, according to science. CRISPR as regulatory arbitrage. How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability?:
One proposal to assess safety is to test-drive autonomous vehicles in real traffic, observe their performance, and make statistical comparisons to human driver performance. This approach is logical, but it is practical? In this report, we calculate the number of miles that would need to be driven to provide clear statistical evidence of autonomous vehicle safety. Given that current traffic fatalities and injuries are rare events compared with vehicle miles traveled, we show that fully autonomous vehicles would have to be driven hundreds of millions of miles and sometimes hundreds of billions of miles to demonstrate their safety in terms of fatalities and injuries. Under even aggressive testing assumptions, existing fleets would take tens and sometimes hundreds of years to drive these miles — an impossible proposition if the aim is to demonstrate performance prior to releasing them for consumer use. Our findings demonstrate that developers of this technology and third-party testers cannot simply drive their way to safety. Instead, they will need to develop innovative methods of demonstrating safety and reliability.
- Benjamin Kuipers explains why, as a Quaker, he refused to accept funding from DARPA, ONR, AFOSR, and ARO: “…it’s possible to have a successful career in computer science without taking military funding”.
- “NSF test finds eliminating deadlines halves number of grant proposals“.
- New, non-technical startup idea: reverse skydiving.
(Seriously. Call me if you want to provide venture capital.) The materialized as the CIA Skyhook. One of the most notable uses was investigating abandoned soviet arctic research stations that were too deep in floating ice to be reached any other way. The system was ultimately abandoned with the arrival of long-range, mid-air refuelable helicopters. - Another story about self-driving cars, this time for truck platoons. I’ve seen articles about this sort of thing for many years, and it’s really hard to tell when actually progress is happening.
- I like this blog post as a way to introduce funnel plots more than the corresponding Wikipedia article. (Just intro material; nothing new for experts.)
- Sunk cost and waiting for the NYC subway, with data.
- The energy efficiency of desalinization is within half an order of magnitude of fundamental thermodynamic limits:
A state-of-the-art facility is now within a factor of two of the theoretical energy minimum, and only 25 percent higher than the realistic minimum for the current reverse osmosis process. In short, it’s going to be tough to squeeze too much more energy out of reverse osmosis, and we’re unlikely to find an alternative method of desalination that will provide a significant boost over that.
But that doesn’t mean that there are no other ways of getting better output for our energy. The total process of desalination turns out to require three to four times the theoretical minimal energy use, since the salt water must be pumped and pretreated, the membranes maintained, and the resulting brine handled afterwards. Some of these things might be amenable to further improvements, and there has been work put into developing membranes that don’t clog up as easily or better pre-filtering of biological materials.
(H/t Carl Shulman. Academic review article.) So you either need to make power cheaper, or you need to take advantage of negentropy leakages during power generation or other industrial processes.
- More from Carl: Going beyond repurposed GPUs, Nvidia will release a dedicated deep-learning (neural-network) processing chip. And on Carl’s blog: charitable lotteries for avoiding economy of scale problems among small-time donors.
- Handwriting recognition for math is getting better and better. (Discussion.) Having given it a brief test run, I think the Apple Pencil is much better than the predecessors I’ve tried. It may be the first handwriting input that is good enough to do this work routinely. Combining these two, along with a Mathematica interface, could finally banish the physicist’s paper notebook.
- It is intuitive to me that certifying the correctness of software often boils down to mathematical theorem proving, but I never thought about the reverse: it is necessary to certify algorithms (by mathematical proof) that are implementing the long computations critical for proving certain theorem (e.g., the Kepler conjecture). From the LEAN theorem prover introduction:
Formal verification involves the use of logical and computational methods to establish claims that are expressed in precise mathematical terms. These can include ordinary mathematical theorems, as well as claims that pieces of hardware or software, network protocols, and mechanical and hybrid systems meet their specifications. In practice, there is not a sharp distinction between verifying a piece of mathematics and verifying the correctness of a system: formal verification requires describing hardware and software systems in mathematical terms, at which point establishing claims as to their correctness becomes a form of theorem proving. Conversely, the proof of a mathematical theorem may require a lengthy computation, in which case verifying the truth of the theorem requires verifying that the computation does what it is supposed to do.
(H/t Dan Selsam.)
- Arithmetic coding is a basic, elegant form of lossless compression.
- Expected cost of influenza pandemic of same order as global warming? I wonder if this is a case of mistakenly comparing pandemic tail risks with global warming bulk risk. (H/t Robin Hanson.)
- A proposal for sending probes to Alpha Centauri. (HN discussion. H/t Peter Woit.) There doesn’t seem to be any stupendous new technical advances as this is based on ideas that have been floating around awhile. This is highly unlikely to materialize, and is mostly notable for the amount of publicity and the big names.
Here’s a good popular article. (H/t 4 Gravitons.) Here are some very vague technical details closely related to the StarWisp proposal by Robert Forward on which it is loosely based.
See also the Japanese solar sail satellite IKAROS and more unhinged speculation.
Here is the list of closest stars and brown dwarfs.
(Brown dwarfs are somewhere between a star and a planet, with just enough mass to sustain deuterium fusion at their core.) It’s mostly just lunchtime speculation, but definitely fun to think about.
- Although your brain is generally fooled by video at 30-60 fps into interpreting the sequence of frames as continuous, your eyes are actually much better at noticing small movements against a background that should be still. This means that augmented reality (AR) displays have to be updated much more rapidly in response to head movement than normal video. What I didn’t realize is that this fast updating could managed by hardware while the software output is updated more slowly (so long, I presume, as the software outputs the full 3D objects):
At this task, the HoloLens is a triumph. The room I was in had various objects pinned to the walls and sitting on tables, and they remained solidly in place throughout my testing. A key part of this is latency; HoloLens apps have to produce 60 frames per second (though this drops to 30fps when recording on the device) to ensure that the application can keep its 3D scene updated to show the right things in the right places.
But the hardware goes further. The mysterious holographic processing unit (HPU), a custom chip found in each device that integrates the accelerometer and spatial data from the device’s sensors to keep track of the world around you, minutely adjusts the display output at 240 frames per second in order to ensure that the 3D objects are drawn exactly where they should be.
And it really works.
Make sure to check out the video.
- The most underrated physicists, according to the brilliant yet infamous Ron Maimon.
- Dan Grover: “More Chinese Mobile UI Trends“.
Links for April 2016
Bookmark the permalink.