Tim Tyler @tim_tyler Twitter profile

Last Seen Profiles

@KevinElevinguy

@StartGameDev_

@abby_finlay

@sakaryader

@gokhantherebel

@davethebluesman

@EffectiveMortga

@GermanGrudny

@JAVICHICAT

@nyanlives

@CirnoFumofumo

@RemingtonPark

@mutuellesinfo

@rokiatraore

@CaoVnQuc2

@JKronbichler

@daeunmeow

@prttyworkshop

@PCtigerfootball

@GetLocoNow

@City_SC

@brasileiraoopre

@Coach_BHarris

@FMGSuite

@DanbiaRodr82167

@r_born22

@definethelimits

@StefanStruve

@zeynep__babacan

@HorribilisPasos

@DrNeilHudson

@MDFzonacaliente

@prince_epenge

@GerkenSascha

@Haruka_mba

@DavidGogina

Tim Tyler

@tim_tyler

7 months

@ylecun A bit of a funny definition of losing: "Top 10 Companies by Market Cap in 2023 #1 Apple Technology $2.728 trillion - #2 Microsoft Technology $2.344 trillion ..."

6

0

26

Tim Tyler

@tim_tyler

7 months

@ylecun I guess the real problem here is that you are cherry-picking the problem area. Google runs on propietary software. Facebook and Instagram are proprietary. Amazon is proprietary. All the main clouds are proprietary. Sure: open source is used in some areas.

6

0

26

Tim Tyler

@tim_tyler

9 months

@iamtrask Re:"Contradicting datapoints are taken as a higher truth than agreement." I think that's just how a standard Bayesian update works. Confirming evidence is weak evidence for a hypothesis - while falsifying evidence is strong evidence against it.

1

0

24

Tim Tyler

@tim_tyler

6 months

@PauseAI @JosephJacks_ @ylecun Dominance is not even a central issue. What happened to: "the AI does not love you, nor does it hate you, and you are made of atoms it can use for something else"...?

1

0

15

Tim Tyler

@tim_tyler

6 months

@_andreamiotti @jasonhausenloy This is the "trust us, we got this" security paradigm.

1

0

15

Tim Tyler

@tim_tyler

7 months

@ylecun @Jake_Browning00 All knowledge is expressible using sequences of 1s and 0s.

6

0

12

Tim Tyler

@tim_tyler

1 year

@albrgr @lukeprog This is not how most regulation works. Usually products are regulated - not hardware or software involved in their implementation.

2

0

12

Tim Tyler

@tim_tyler

8 months

@RokoMijic Consensus in evo. biology is that there are typically many causes of aging - in large creatures. If one mechanism killed you faster than the others, then it would attract maintenance resources. The result is many different things conspiring to kill you at around the same time.

5

0

12

Tim Tyler

@tim_tyler

1 year

@RespectToX @d_feldman The first hit is often free.

0

10

Tim Tyler

@tim_tyler

6 months

Free speech for chatbots.

3

2

10

Tim Tyler

@tim_tyler

1 year

"Human-level AI" is a dumb concept. By the time machines are at least "human-level" at 50% of the things humans often do, they will be a million times better at other things. Since it's a poor-quality concept, it's best to ditch it before it further messes up thinking and debates

0

8

Tim Tyler

@tim_tyler

1 year

Who benefits from doomer alarmism? Probably current incumbents - since it promotes overregulation, and so raises the barrier for entry.

2

0

9

Tim Tyler

@tim_tyler

7 months

@romanyam Regarding comparisons with fire and language - I would go back further. We are likely to be facing a genetic takeover. The last one of those was around 4 billion years ago.

2

0

5

Tim Tyler

@tim_tyler

10 months

Imagine an old cracked dam. It's going to fail. Your must protect those downstream. You can block holes (to buy time) or enlarge holes (to allow water through and reduce pressure on the dam). Perhaps this second strategy is neglected because it seems more risky and less obvious.

0

2

9

Tim Tyler

@tim_tyler

8 months

@RichardSSutton This seems like the "lie down and close your eyes" plan.

0

7

Tim Tyler

@tim_tyler

1 year

Mitigating the risk of extinction from ignorance and stupidity should also be a global priority.

0

7

Tim Tyler

@tim_tyler

1 year

@ylecun @per_arneng You don't know how to "guarantee AI safety" either.

1

0

8

Tim Tyler

@tim_tyler

1 year

Attempting to ban things that people want has been attempted before. Alcohol prohibition, prostitution and drugs for example. It creates black markets which are run by criminals and are difficult to regulate.

1

7

Tim Tyler

@tim_tyler

1 year

@DrJimFan Or maybe not:

Hyena Hierarchy: Towards Larger Convolutional Language Models

Recent advances in deep learning have relied heavily on the use of large Transformers due to their ability to learn at scale. However, the core building block of Transformers, the attention...

arxiv.org

0

7

Tim Tyler

@tim_tyler

1 year

@leecronin This one can be looked up:

Emergence - Wikipedia

en.wikipedia.org

0

6

Tim Tyler

@tim_tyler

7 months

@robinhanson This reminds me of the climate change concerns. We're going to get machine superintelligence this century. That dwarfs these other minor concerns.

1

0

7

Tim Tyler

@tim_tyler

1 year

@robinhanson The quote is also nuts: "members of the same race, relative to the species as a whole, are related to one another (rH = 0.18–0.26) almost as closely as half-siblings (rH = 0.25)." Surely that indicates that they are mixing together different ways of measuring of relatedness.

0

6

Tim Tyler

@tim_tyler

1 year

@PostOpinions @MaxBoot UK and US promised to defend Ukraine in the Budapest agreement - in exchange for nuclear disarmament. It seems as though they are not doing a good job.

3

1

6

Tim Tyler

@tim_tyler

5 months

@MatthewJBar The improvements over GPT-4 are all in an unreleased, vaporware product that we might get to see sometime next year.

0

7

Tim Tyler

@tim_tyler

11 months

The "p(doom)" term has gone viral. I think I may have come up with it. I've been using it since at least 2009 - with numerous of references in 2010:

This Failing Earth — LessWrong

Suppose I told you about a certain country, somewhere in the world, in which some of the cities have degenerated into gang rule. Some such cities ar…

www.lesswrong.com

4

0

7

Tim Tyler

@tim_tyler

10 months

@MikeAnissimov2 Counter-example: trade.

1

0

6

Tim Tyler

@tim_tyler

1 year

@robinhanson Or maybe the respondents accounted for evolutionary convergence.

1

0

7

Tim Tyler

@tim_tyler

9 months

@DrTechlash "I spoke to a few of those who signed the letter, and it was clear that they did not all agree entirely with everything it said" - that sounds fairly normal and doesn't seem like much of a news story.

0

7

Tim Tyler

@tim_tyler

7 years

The big bang is the belly button of the universe.

0

5

Tim Tyler

@tim_tyler

9 months

@realGeorgeHotz Computation doesn't depend on overwriting bits. For example, see reversible computation. The Landauer limit is irrelevant to reversible computation. Moore's law is different - since it is to do with dollars.

0

Tim Tyler

@tim_tyler

9 months

@floates0x As many have previously noted, it is pretty easy to make a video just like this with ordinary magnets.

4

0

7

Tim Tyler

@tim_tyler

1 year

@Plinz As machine capabilities grow, first they complement humans, then they substitute for humans. The humans probably won't be unemployed - but wages could fall to what you would pay a machine.

1

0

6

Tim Tyler

@tim_tyler

1 year

Some claim that we only get one try at building superintelligence. Apparently, if we fail, we don't get to try again, because we are already dead. We are not allowed to make any civilization-destroying mistakes, but that's always true, and other mistakes seem permissable.

1

0

6

Tim Tyler

@tim_tyler

11 months

@evanrmurphy @pmddomingos @AndrewYNg @geoffreyhinton That's what he said. It is probably safe to assume that that's what he meant.

0

7

Tim Tyler

@tim_tyler

1 year

@RokoMijic Most forms of therapy are broken on purpose - through rewarding failure and cultivating dependency.

0

6

Tim Tyler

@tim_tyler

1 year

@Plinz Decensoring:

MMitchell

@mmitchell_ai

1 year

Just read the draft Generative AI guidelines that China dropped last week. If anything like this ends up becoming law, the US argument that we should tiptoe around regulation 'cos China will beat us will officially become hogwash. Here are some things that stood out. 🧵

27

186

736

1

0

5

Tim Tyler

@tim_tyler

9 months

@AdamBRozycki @mustafasuleyman It seems to be: what if the bad guys get hold of it? As if the bad guys were going to respect the terms of the software license in the first place. If license restrictions won't work, keeping the technology secret is sometimes proposed. The "trust us" model.

2

0

4

Tim Tyler

@tim_tyler

7 months

@octonion @leecronin @RosieRedfield @Nature The paper does not claim the planets like Earth are closed systems in the first place.

2

0

4

Tim Tyler

@tim_tyler

9 months

@lVlarty @AISafetyMemes @tszzl Destruction has a kind of thermodynamic advantage over construction. They saying goes "it is easier to destroy than create". However, as you say, this is missing some factors from the accounting - and if you include them, the conclusion is often reversed.

1

0

6

Tim Tyler

@tim_tyler

7 months

@romanyam Am still uncomfortable with the "we have to get it right on the first try" rhetoric. We have been trying and failing since the 1950s - and there will be plenty more failures in the future. Severe, unrecoverable setbacks are possible, but they aren't the only kind of failure.

0

1

Tim Tyler

@tim_tyler

1 year

@Brewgaloo_ Retroanglofuturism.

0

6

Tim Tyler

@tim_tyler

9 months

@janleike Open source LLMs can reliably survive and spread today - with the help of humans. The human in the loop slows them down a bit and tthe human has its own preferences - but it's not a big deal. As it turns out, autonomy is not worth all that much.

0

Tim Tyler

@tim_tyler

1 year

@leecronin Invalid syllogism alert.

0

6

Tim Tyler

@tim_tyler

6 months

@michaelhunterAU @geoffreyirving On what basis?

1

0

5

Tim Tyler

@tim_tyler

7 months

@sashachapin That's not a real quote from Steve Jobs.

0

2

Tim Tyler

@tim_tyler

1 year

Listening to the doomers, I don't think many of them realize how bad their chances are with no machines. 99% of all species go extinct. The odds for large vertebrates are especially bad. Machines are our most likely route to salvation.

0

4

Tim Tyler

@tim_tyler

6 months

@ylecun The "AI doomer movement he jumpstarted"...? Reading the "Superintelligence" book, it seems as though most of Bostrom's points came from Yudkowsky.

0

4

Tim Tyler

@tim_tyler

1 year

@MatthewJBar IMO, the basic idea is right, but the timescale is more like decades. Short timescales were probably cited (e.g in the "RSI" essay) because short timescales are scary and so stimulate donations.

1

0

5

Tim Tyler

@tim_tyler

6 months

@McaleerStephen It's a sneaky way to draw attention to your paper.

0

4

Tim Tyler

@tim_tyler

8 months

@karpathy "Love is all you need" - The Beatles -

Love Is All You Need - Beatles

A tribute to this wonderful band...

www.youtube.com

0

Tim Tyler

@tim_tyler

1 year

Is there any difference from the industrial revolution and the chance of destroying civilization with nuclear weapons? If so, it seems as though it is mainly because more powerful technology is involved.

1

0

4

Tim Tyler

@tim_tyler

10 months

@togelius @dioscuri @davidchalmers42 That is silly. Fire is "self amplifying". Life is "self amplifying". Life is not a simple system.

0

5

Tim Tyler

@tim_tyler

5 months

@TheStudyofWar @nataliabugayova The US is widely considered to have lost the Vietnam War. The US military is impressive - but distant international conflicts are always on a budget.

5

0

8

Tim Tyler

@tim_tyler

1 year

A meta-learning perspective suggests that back propagation is probably a dumb strategy. If you had an intelligent machine that understood your network to help update your weights then it could do much better. This may sound circular - but really it's more like a wheel.

0

3

Tim Tyler

@tim_tyler

13 years

Internet killed the copyright law.

0

1

0

Tim Tyler

@tim_tyler

9 months

@IPCC_CH That's a funny-looking projection.

0

Tim Tyler

@tim_tyler

7 months

@sashachapin What do you have against AI waifus?

0

Tim Tyler

@tim_tyler

1 year

@RokoMijic Unemployment rate in the USA is 3.4 percent - a 50 year low.

1

0

5

Tim Tyler

@tim_tyler

1 year

Banning things that many people want creates black markets run by criminals that are difficult to regulate. Look at what happened with prohibition, prostitution and drugs. The Prison Industrial Complex is the main beneficiary.

0

4

Tim Tyler

@tim_tyler

7 months

@RokoMijic So: maybe all the stuff about "misaligned" machines is an irrelevant distraction - and most of the problem is with what we already have: corporations, governments, humans and technology.

3

0

5

Tim Tyler

@tim_tyler

1 year

@ravisparikh Note that "now regrets his life’s work" is a quote from Cade Metz, not Geoffrey Hinton. It is not terribly clear what Hinton said on the topic.

1

0

5

Tim Tyler

@tim_tyler

7 months

@romanyam Assuming that superintelligence is uncontrollable, what are the options for preservation of our values? Here is what I can see: relinquishment, value transfer, merger and preservation via instrumental goals.

1

0

Tim Tyler

@tim_tyler

7 months

@RokoMijic Backpropagation is a dumb strategy. I mean: your network would learn faster and better if there was an expert system adjusting its weights in response to an error function - instead of a mindless "gradient descent" process.

1

0

5

Tim Tyler

@tim_tyler

1 year

@bengoertzel @SingularityNET Re: "If you believe, as we do, that at some point, AI - AGI - is going to be extremely, unbelievably potent, then it just does not make sense to open-source. It is a bad idea..." - so: let's keep it to the mega-corps and make sure no one else has the tech to defend against them.

0

5

Tim Tyler

@tim_tyler

6 months

@Simeon_Cps @hokiepoke1 @ThierryBreton @vestager @IoanDragosT @brandobenifei @LucillaSioli What? That doesn't make sense.

0

5

Tim Tyler

@tim_tyler

7 months

@sashachapin Why do you think AI has been bad so far? I mean apart from your examples - which could be cherry picked or mistaken - did you attempt any sort of cost-benefit analysis?

0

Tim Tyler

@tim_tyler

9 months

@IPCC_CH Projecting out to 2100 without huge error bars means you don't think machine superintelligence is going to be a big deal.

0

Tim Tyler

@tim_tyler

1 year

Absurd beliefs actually signal group membership better. They are costly (see costly signalling theory) and difficult for outsiders to accurately fake.

ExiledInfoHaz

@ExiledInfoHaz

1 year

In other words, irrationality isn't a series of random mistakes and glitches. It's actually an adaptation where people bend their factual beliefs to be more like those of the group, even if they are fairly obviously wrong.

4

1

37

0

3

Tim Tyler

@tim_tyler

6 months

@ylecun The reason LLaMa 2 isn't open is because it doesn't have an open source license and instead places restrictions on commercial use.

0

3

Tim Tyler

@tim_tyler

5 months

@CFGeek @MatthewJBar It is not out of the door yet.

0

4

Tim Tyler

@tim_tyler

5 months

@keith_dorschner @mattwridley @DrJBhattacharya In science there is always doubt. Absence of doubt is faith. That means you can't update on evidence.

0

3

Tim Tyler

@tim_tyler

6 months

@GaryMarcus @NLeseul @sebkrier Usually, the people keeping the secrets *are* the bad actors. The reason they are keeping things secret is because they have something to hide that they don't want others to know about.

0

1

3

Tim Tyler

@tim_tyler

1 year

@RokoMijic "Eugenics" is conflated with "negative eugenics" in most people's minds. It thus suffers from guilt by association. As a result, few men explicitly endorse it either.

1

4

Tim Tyler

@tim_tyler

6 months

@liron @gdb @sama 1 is GPT-4 while 2 is GPT-5.

1

0

3

Tim Tyler

@tim_tyler

6 months

@DrNikkiTeran Why did they not control using a search engine? Is that because doing so would invalidate their conclusions?

0

4

Tim Tyler

@tim_tyler

11 months

@Roko__eth It only works if both sides can agree on what the outcome would be. In the example given, both sides thought they could win.

1

0

4

Tim Tyler

@tim_tyler

6 months

@AlecStapp The cop explained that this is conventionally called:

Bisection method - Wikipedia

en.wikipedia.org

0

4

Tim Tyler

@tim_tyler

9 months

@primalpoly @liron @janleike @OpenAI The premise seems dubious, but working for OpenAI would make sense for a moral angel if they believed that OpenAI had a better chance of success than its rivals - and that by working for them there would be a chance to positively influence the outcome.

2

0

3

Tim Tyler

@tim_tyler

5 months

@MikePFrank @satyanadella Then everyone else will want a piece of the action.

0

4

Tim Tyler

@tim_tyler

1 year

@PeterDClack @lalithramachan1 Denial of data.

2

0

3

Tim Tyler

@tim_tyler

6 months

@tegmark It seems odd to imagine machine superintelligence using very much solar power.

2

0

4

Tim Tyler

@tim_tyler

7 months

@RokoMijic @MikePFrank @gcolbourn

Revisiting Algorithmic Progress

We use a dataset of over a hundred computer vision models from the last decade to investigate how better algorithms and architectures have enabled researchers to use compute and data more efficient...

epochai.org

2

0

3

Tim Tyler

@tim_tyler

7 months

@octonion @leecronin @RosieRedfield @Nature Why you are talking about "closed systems" is not clear. That's not mentioned. Perhaps you are projecting.

1

0

4

Tim Tyler

@tim_tyler

6 months

@liron @gdb @sama Greg also says: "I think GPT-5 will just be different in some way that’s hard to describe now." I think this disagreement is mostly fabricated.

0

3

Tim Tyler

@tim_tyler

7 months

@leecronin @Nature Universal Darwinism unites physics and selection without "assemblies".

1

0

2

Tim Tyler

@tim_tyler

11 months

@liron @pmarca This makes Marc look like quite a fruitcake. It looks as though it comes from this video, dated two days ago:

AI Will Save The World with Marc Andreessen and Martin Casado

This week, a16z’s own cofounder Marc Andreessen published a nearly 7,000-word article that aimed to dispel fears over AI's risks to our humanity – both real ...

www.youtube.com

0

1

4

Tim Tyler

@tim_tyler

1 year

@jesswhittles Different timescales too. We will likely see the benefits before we get to any of the more serious risks.

1

4

Tim Tyler

@tim_tyler

1 year

@sama If they are the same, then why are there two names?

12

1

5

Tim Tyler

@tim_tyler

7 months

@ATabarrok In the beginning there were vi and emacs...

0

3

Tim Tyler

@tim_tyler

9 months

@Plinz Re: "Let states create their own FDAs" - competition might help, but this also seems like duplicated effort.

1

0

3

Tim Tyler

@tim_tyler

1 year

Open AI seem to have fixed a good many of the security exploits that led to their product giving advice about dangerous or illegal activities. Hopefully no bunnies got hurt in the interim.

1

0

2

Tim Tyler

@tim_tyler

1 year

@RokoMijic I'm pretty sure that that contradicts the facts. Plenty of smart people are not terrified.

0

2

Tim Tyler

@tim_tyler

8 months

@Aella_Girl I've tried the "no-poo" approach in the past. My testimony is that "it works" - but then so does the "poo" approach. I wouldn't mind seeing some stats on split ends, length, dandruff, baldness, greyness, etc.

0

1

Tim Tyler

@tim_tyler

6 months

@RokoMijic @primalpoly @BogdanIonutCir2 @robertskmiles I think we can agree on that then. 2,000 years is very small compared to the likely timescale of alien contact. It's much more important to avoid making a mess of things right now than to go pedal-to-the-metal to keep up with hypothetical future encounters with advanced aliens.

0

3

Tim Tyler

@tim_tyler

1 year

The Singularity is Here - David Perell

The Singularity is already here, and it began in Western Europe at the end of the 16th century.

perell.com

2

0

2

Tim Tyler

@tim_tyler

7 months

@MikePFrank 3 billion years to make a cave man. 3 million more years to make Albert Einstein and John von Neumann. The last bit could come in a rush.

1

0

3

Tim Tyler

@tim_tyler

1 year

OK, doomer: you can put those missiles down now.

0

2

Tim Tyler

@tim_tyler

9 months

@liron The size of the human brain has been gradually decreasing for the past 28,000 to 34,000 years. Humans have been domesticated by their own organizations and institutions. Peak encephalization is in the past.

0

3

Tim Tyler

@tim_tyler

5 months

@jpsenescence @EleanorSheekey I read the article. I says: "Clearly, there is a software program, encoded in the DNA, that is far more advanced, with much greater algorithmic complexity, than any computer program." What? 3 gigabytes?

2

0

2

Tim Tyler

@tim_tyler

11 months

@Faceplants00 @JohnCav60020747 @Kevin_McKernan @BretWeinstein "Failing fast" - as many engineers call it.

1

0

3