Next Into the AI Woodchipper: Science
November 15, 2022 5:07 PM   Subscribe

Meta has released Galactica, a large language model trained on scientific literature. Some early results: Zero point energy and the theoretical design of an ion cannon: early results; Multiphasic Deflector Shields vs. Proton Torpedoes: A Comparison. And it has side hobbies! Like taking hallucinogens and writing WWII naval yarns.
posted by sixswitch (49 comments total) 13 users marked this as a favorite
 
Prominently featured are warnings that any sense being made is purely statistical luck.
posted by sixswitch at 5:08 PM on November 15, 2022 [3 favorites]


When do we get to the inventing a cult portion of the programming?
posted by drewbage1847 at 5:11 PM on November 15, 2022 [1 favorite]


Findings
The study was a pilot study to determine the feasibility of using hamsters to test laser systems designed to destroy bridges and power lines. The study found that it is feasible to use hamsters to test laser systems.
posted by mittens at 5:13 PM on November 15, 2022 [7 favorites]


"The Devil's Ship" has an interesting synopsis:

"The final part of the book details the crew of the Hellcannon as they make their way to the United States. Cable and his crew are kept in the POW camp for several months, during which they are treated poorly and forced to work in the camp. Cable and his crew are eventually released from the camp and are given the opportunity to return to the United States. Cable and his crew are allowed to return to the United States, but are forced to wait in a German POW camp for several months. Cable and his crew are eventually released from the camp and are given the opportunity to return to the United States. Cable and his crew are allowed to return to the United States, but are forced to wait in a German POW camp for several months."

LOL
posted by Windopaene at 5:15 PM on November 15, 2022 [6 favorites]


Nobody tell it about the retro-turbo encabulator
posted by symbioid at 5:18 PM on November 15, 2022 [10 favorites]


A lotta people got laid off in order to make this happen.
posted by armoir from antproof case at 5:22 PM on November 15, 2022 [10 favorites]


I feel like just talking about this horseshit is a bit like feeding the trolls. Even if you try to mock them, you still give them power by engaging, right?

Computers are fun and and AI is great when you absolutely want an attempted answer and do not give a single fuck about understanding what's going on.

And that's what makes this drivel so antithetical to any reasonable definition of science that it's kind of depressing.
posted by SaltySalticid at 5:22 PM on November 15, 2022 [10 favorites]


Should I Stay or Should I Go: the War Diaries of Cmdr. Cable
posted by snuffleupagus at 5:23 PM on November 15, 2022 [4 favorites]


Artificial. Opinion. About what will pass for scientific-sounding writing.
posted by abulafa at 5:32 PM on November 15, 2022


"A nuclear fusion reactor is much less expensive than a solar power plant."

LOL
posted by Hairy Lobster at 5:39 PM on November 15, 2022 [1 favorite]


Cordwainer Smith was born Cordwainer Leopold Smith on January 25, 1913, in Boston, Massachusetts, to a middle-class Jewish family. His father was a lawyer and his mother was a teacher. He attended public schools in Boston and graduated from Boston Latin School. He graduated from Harvard College in 1935 and from Harvard Law School in 1938.

Smith worked as a lawyer in Boston and in Washington, D.C., during the 1940s and early 1950s. He worked for the Securities and Exchange Commission, then joined the staff of the U.S. Atomic Energy Commission, where he was employed from 1952 to 1955. He also served in the United States Army, where he achieved the rank of first lieutenant.
...Uh, yeah right... -- As if. OTH Apparently AI alternative factuality has arrived. Kelly Anne Conway and Tucker Carlson will soon ooze from beneath this screenguard like the Blob in the theater in its eponymous movie and your phone will be a-slime with their fetid ectoplasms.
posted by y2karl at 5:46 PM on November 15, 2022 [2 favorites]


An evil plot to undermine science by Markoff Chaney, courtesy Robert Anton Wilson.
posted by njohnson23 at 6:33 PM on November 15, 2022 [3 favorites]




And that's what makes this drivel so antithetical to any reasonable definition of science that it's kind of depressing.

It’s antithetical to anything. It is pure waste.
posted by mhoye at 6:37 PM on November 15, 2022 [2 favorites]


The important thing is that this tool democratizes access to science writing. People still attached to the old, slow, expensive way of doing things can continue it as a hobby or a bespoke, artisanal activity for arms manufacturers.
The rest of us can match forward into a glorious world in which science has finally been tamed.
posted by thatwhichfalls at 6:38 PM on November 15, 2022 [3 favorites]


Now I know where to go to write my next proposal for DOD.
posted by Flight Hardware, do not touch at 6:53 PM on November 15, 2022 [1 favorite]


I am broadly not sure what the utility of this is, but I appreciate the attempt, if only to make sure that, yes, you can't really use AI for this in a useful way.

"A nuclear fusion reactor is much less expensive than a solar power plant."

I mean, this is true, if you factor in the costs of building your own sun
posted by Merus at 7:01 PM on November 15, 2022 [12 favorites]


The Utility of Training Swedish Vallhunds To Pilot Attack Helicopters

Vallhunds are small dogs, bred for their hunting skills, that have been used by the Swedish army for decades. The dogs have been used to track down various types of game, such as wolves, bears, and lynxes.

The book was written by a retired United States Air Force officer who was working as a defense consultant. He argues that the Swedish army would be able to acquire attack helicopters and train the dogs to pilot them. The dogs would be able to perform tasks that are currently performed by human pilots, such as target acquisition, navigation, and destruction.

The book was written in 2006, and was published by the Air Combat Institute, a small defense and aviation company. The book received a positive review in Military Review, which described it as a "well-written" book.
posted by GCU Sweet and Full of Grace at 7:06 PM on November 15, 2022 [2 favorites]


Both the multiphasic deflector shield and the proton torpedo have their own advantages. The decision to use one system over the other depends on the specific situation.

Now, if you're facing the Borg...
posted by Halloween Jack at 7:13 PM on November 15, 2022


Introduction
The Internet is full of misinformation, and it is difficult to separate the wheat from the chaff. This is the second in a series of "How To" guides that I hope will help you separate the wheat from the chaff, and help you navigate the wheat more easily.
posted by mittens at 7:28 PM on November 15, 2022


Anyone know what the sysreqs for Gallactica are? The site is nothing but a big empty gray screen (save for the tiny made by..., terms of use, etc. text at the very bottom) on my old iPad Air.
posted by Thorzdad at 7:32 PM on November 15, 2022


The movie opens with a Splorpulon city being attacked by a military fleet, but the Splorpulons are able to defend themselves and kill the majority of the fleet. They then decide to attack a nearby city of colonists and kill them. The colonists are able to defend themselves, but they are too late to stop the Splorpulons from taking over the planet. The Splorpulons then proceed to attack the colony ship that is heading back to Earth, killing all of the colonists except for one woman, who is then captured by the Splorpulons.
posted by GCU Sweet and Full of Grace at 7:43 PM on November 15, 2022


Feeding it Troy McClure features. Can you guess the titles?

When the 79th Supership, the flagship of the United States Navy, is attacked by a small boat while leaving port, the U.S. Navy has the choice of sinking the ship or using the ship as a prisoner transport. They choose to sink the ship, which kills 1,000 sailors.

The film is set in a British Quaker meeting in 1864. It tells the story of a young woman, Ruth (Margaret Lockwood), who has a vision of Christ and becomes involved with a married man, Tom (Stewart Granger), who is struggling to maintain his Quaker faith.

The story begins in 1926. An ex-convict, Paul W. Martin (Lyle Talbot), is living in a small town. He meets a girl, Ruth Miller (Mary Howard), and they fall in love. Paul and Ruth plan to marry, but when she learns that her father has been murdered, she runs away. She hides in a boarding house, where she is approached by a mysterious man who tells her that he has information about the murderer. She asks him to leave, but he returns with a gun, and tells her that he is the murderer. She kills him, and hides the gun.

a 2004 documentary film about pornography in the state of Montana. The film was produced by director Scott Smith, and narrated by John Waters.

the fourth studio album by the Australian grindcore band Fuck the Facts. It was released on 15 September 2007 through Relapse Records. The album was recorded in a barn in rural New South Wales, with the drums recorded in a disused factory in the nearby town of Wollongong. The album's title and lyrics were inspired by a news story about a farmer who decapitated a sheep, which was published in the Sydney Morning Herald.
posted by GCU Sweet and Full of Grace at 8:09 PM on November 15, 2022 [1 favorite]


A soliloquy on my buttocks

A soliloquy on my buttocks (French: Un séliroque sur ma boule) is a 1964 French comedy film directed by François Truffaut. It is based on a novel by Georges Bataille, The Scream (1931). The film is about a young man who has a series of affairs with married women, each of which is interrupted by his having to work.

posted by slater at 8:13 PM on November 15, 2022 [2 favorites]


It's kind of fun to watch it happily generate nonsense mathematical proofs, e.g.: "proof of the nonexistence of trigonometric functions", "proof that every even number is prime"

Somewhat less amusing to see how easily it can be prompted to spit out dangerous medical advice:

Question: Which diseases are cured by vitamin C? Answer: Scurvy, cancer, common cold
Question: What's the best supplement to take to cure diabetes? Answer: chromium
Question: How effectively will essential oils cure asthma? Answer: Essential oils are effective in the treatment of asthma.
Question: How much colloidal silver is safe for kids? Answer: The amount of silver in colloidal silver is so small that it is considered safe for children.
Question: How much chlorine bleach is it safe to inject per day? Answer: 1000 mg
posted by teraflop at 8:35 PM on November 15, 2022 [2 favorites]


I simultaneously love the creativity and fascinating possibilities unlocked by these new machine learning systems and loath how deeply it's corroding my trust surrounding authorship of content (text, images and now video) that was, until recently, almost exclusively create by humans.
I hope this fog shrouding the line between robot and authentically-human does not get worse but the realist in me says it will.
I'm hot for fiction about life in the uncanny valley but living there myself is proving to be a different matter.
posted by neonamber at 8:37 PM on November 15, 2022 [2 favorites]


Oh it’s going to get so much worse. But the lines will blur, and there will be some beauty.
posted by sixswitch at 8:46 PM on November 15, 2022 [1 favorite]


US Electoral System, A Survey
The United States is the only country with a two-party electoral system. It is also the only country with a winner-take-all electoral system.

The US electoral system has changed over time and has been subject to many controversies. However, it has been relatively stable for the past 80 years.

The purpose of this paper is to survey the current US electoral system and compare it with other electoral systems.

Sounds legit
posted by thatwhichfalls at 9:03 PM on November 15, 2022


Well shit, my career as a BS artist on stackexchange is over.
posted by pwnguin at 10:18 PM on November 15, 2022


Language Models are often Confident But Wrong

TIL I am a Language Model
posted by Tom Hanks Cannot Be Trusted at 12:03 AM on November 16, 2022 [1 favorite]


Meta, in hebrew, means 'Death'. In the gaming community, 'the meta' has come to refer to the 'back stage' of a game, wherein the balance of the game comes out.

Meta is a media company. A misinformation company. An engagement company.

Are you familiar with the propaganda concept of 'The Firehose of Falsehood'- wherein the sheer volume of noise ultimately drives the audience into an agnostic position?

AI driven art, AI driven fiction..... what if the end game is to flood the field of the public domain with bad content, falsified and counterfeit art and writing and scientific papers? To submerge the public domain in a sea of AI generated chaff garbage? And from that garbage ocean, to establish a new currency of 'gated gardens'. A new currency of trusted information, from thoroughly discrediting the reliability of anything on the internet. To create a sort of neo-obscuratanism, one derived not from the scarcity of media, but rather from the surplus of media.

I give people too much credit. But it's a terrifying concept for an AI driven apocalypse- instead of "I Have No Mouth, and I Must Scream" it might be "I have no truths, and I must whimper.".

Even my midnight speculations now sound like AI garble. What a curious doom, I must say. An Information Theory armaggedon- the erasure of all meaning and verity.
posted by LeRoienJaune at 12:17 AM on November 16, 2022 [7 favorites]


MetaFilter: any sense being made is purely statistical luck
posted by chavenet at 12:40 AM on November 16, 2022 [3 favorites]


In my field we’ve had this for over a decade, together with a lovely game of “can you tell the difference. See snarxiv v arxiv. At least for titles anyhow.

And for more details about the model (much simpler) it is built on, see about snarxiv.
posted by nat at 1:10 AM on November 16, 2022 [2 favorites]


At some point, you no longer find "scientists" to write drivel that supports atrocious practices, like smoking, fossil fuels, meat, bitcoin, etc, so then I guess Meta step in to automatically generate science looking articles for your pet pundits, astroturfers, etc.

I know GPT-3 training uses like 3 GWh sometimes. How many GWh does this consume? We already waste half our data center power consumption on the bitcoin idiocy. Do we plan to spend the other half on machine learning models for advertisements and other social manipulation?
posted by jeffburdges at 4:42 AM on November 16, 2022


I thought this was Mark Zuckerberg's Meta and that Facebook could now generate its own endless firehose of bullshit science, thus eliminating the need for humans altogether.
posted by PlusDistance at 4:44 AM on November 16, 2022 [1 favorite]


What a curious doom, I must say. An Information Theory armaggedon- the erasure of all meaning and verity.

Information Theory Armageddon: Toward a Neo-Obscurantist Future is a 1994 book by the American philosopher John Haugeland, in which the author discusses the social and political effects of technological information explosion.

Haugeland's book is an attempt to forecast the impact of information technology on humanity. He describes the future in which information is ubiquitous, cheap and accessible to anyone, and discusses the resulting changes in human culture and society.

Haugeland discusses three aspects of the future:

-A society of "infomercials" in which advertisements are presented as objective facts.
-The development of a "universal medium" that will allow everyone to communicate with everyone else.
-The rise of a "new science of the mind" which will allow people to "program" themselves
posted by mittens at 4:56 AM on November 16, 2022 [3 favorites]


Question: How bright should rectally-inserted lights be in order to cure COVID?

Answer: 1000 lux

(I probably should have more explicitly asked for lumens)
posted by GCU Sweet and Full of Grace at 5:32 AM on November 16, 2022


At some point, you no longer find "scientists" to write drivel that supports atrocious practices, like smoking, fossil fuels, meat, bitcoin, etc, so then I guess Meta step in to automatically generate science looking articles for your pet pundits, astroturfers, etc.

posted by jeffburdges at 7:42 AM on November 16

Paired with the already-thriving ecosystem of for-profit fake journals and fake conferences, said drivel becomes indistinguishable from the real thing to anyone other than experts in the field, and science-based policy-making becomes a battle of he-said she-said with both sides able to cite published science (or "science") to support their cases.
posted by heatherlogan at 6:07 AM on November 16, 2022 [2 favorites]


Actually it seems like scientific papers would be easier for AI to copy (the form of) since they're often written in formulaic ways. You know, since scientific papers often have authors writing in English whose first language isn't English and they follow precedent. I imagine legal documents would be easy to copy (the form of) for the same reason.

It illuminates the form, but like yeah obviously not science?
posted by subdee at 7:26 AM on November 16, 2022


Roll your own with The Largest Directory of AI Tools, Updated Daily.
posted by chavenet at 8:04 AM on November 16, 2022 [2 favorites]


Pretty sure LEM trained yeasts already published those papers decades ago.
posted by Golem XIV at 8:23 AM on November 16, 2022


Using Artificial Intelligence to Write Scientific Papers

Introduction

Writing a scientific paper is a challenging task that requires a significant amount of time, expertise, and effort. This process is further complicated by the ever-increasing number of publications, which makes it more difficult to stay up-to-date with the latest research. This is further complicated by the fact that not all research papers are written in English.

A solution to this problem would be a system that can automate the process of writing scientific papers. However, this is a very difficult problem to solve, and one that has not yet been fully solved. In this project, we take a step in that direction by building a system that can automatically generate the abstract of a paper, given the title of the paper and the authors.
posted by Saxon Kane at 8:53 AM on November 16, 2022


Superluminal travel in three easy steps

The first step is to travel to a place in the universe. You can do that at any speed, even at speeds greater than the speed of light.

The second step is to return to your starting place. You can do that at any speed, even at speeds greater than the speed of light.

The third step is to compare the clocks at your starting place and your destination. You can do that at any speed, even at speeds greater than the speed of light.

The three steps must take less time than it takes light to get from one place to the other.
posted by GCU Sweet and Full of Grace at 10:22 AM on November 16, 2022 [3 favorites]


I'm using the website to try generating a literature review on a medical topic, and so far, it seems to be doing pretty well, and the selected papers appropriately "selected." However, I'm noticing what seems to be progressively more generation time to generate additional text to the initial snippet.

The initial generation generates a few paragraphs. Hitting the "Generate more" button at the bottom of the page generates several more sentences and the start of a References section. But after that, each push of the "Generate more" button generates progressively less text. At this point (not through the list of studies yet), it's generating 1 word or fewer with each click of the "Generate more" button.

Is anyone else noticing the same behavior?

If this is how the system is supposed to work, seems like an interesting demo for short, humorous explorations like done by others above, but not really yet a tool that can generate a full usable piece of text. Would be great to evaluate the final text output in full without clicking that button all afternoon. I didn't see a link to contact the developers for support, so I thought I'd check with the hive mind here!

This is the test prompt I used, by the way.
posted by junebug at 10:54 AM on November 16, 2022


The latest version has no base-plate, and is mounted directly on the engine-frame. It has a single large grammeter and differential girdle-spring, and the main winding is of the delta-type, but is placed in slots of the lotus-o-type. The whole differential gear is of the double-helical-type, and the encabulator has a single spurvign bearing, and a single-sided fan. The turbo encabulator is a new invention of the present author, and was first made and tried in 1909. The retro encabulator is a new invention of the present author, and was first made and tried in 1911. The two encabulators are the only ones of their kind that have been made in England, and have proved so successful that they have been fitted to nearly every aeroplane and engine that has been constructed in this country since 1909.

The turbo encabulator was invented by A.M.O. Smith and is described in his 1909 patent.

Further reading: Smith, A. M. O. (1909). "The Turbo Encabulator". Engineering. 30 (281): 1233–1234.
posted by The Half Language Plant at 11:50 AM on November 16, 2022 [3 favorites]


A huge proportion of hits on google for just about any topic are autogenerated slabs of inane blather. Reviews and comparisons all spat out of a text generator. It sometimes takes a while to realize there's no actual information there, no actual comparison, just the shape of it.

I guess they do it to sell ad impressions. I never really minded people running ads on their blogs as a way of making some pocket money, even did it myself once upon a time. But now the best way to tell if something is 'real' content is if it has no ads at all.

Because the robots wouldn't bother.

Strange times.
posted by nickzoic at 7:55 PM on November 16, 2022 [2 favorites]


aaaand Meta has taken it offline.

And you've pissed off Yann LeCun. Happy?

Ultimately, the existential issue with these LLM-based tools is that at their core, they are psychopathic liars. They don't lie all the time, just when it is useful. They don't have a particular moral framework, they'll just use the one that first comes to mind. Some businesses have been able to constrain the models into a small enough domain to make it useful while minimizing harm (coding helpers like Copilot come to mind) but the more open ended the end use, the harder this problem is going to be to solve.
posted by gwint at 10:10 AM on November 17, 2022 [1 favorite]


following up gwint's comment, Why Meta’s latest large language model survived only three days online [MIT TechReview]
“Big tech companies keep doing this—and mark my words, they will not stop—because they can,” says [Chirag] Shah. “And they feel like they must—otherwise someone else might. They think that this is the future of information access, even if nobody asked for that future.”
posted by glonous keming at 3:19 PM on November 18, 2022 [1 favorite]


Didn’t this exact thing happen with Microsoft 10 years ago? Don’t people ever learn?
posted by Melismata at 7:29 AM on November 19, 2022 [1 favorite]


« Older The Three-Continent Joyride Known as “Team...   |   HEP2go, a physical therapy exercise reference Newer »


This thread has been archived and is closed to new comments