Shane Legg
@ShaneLegg
Followers
54K
Following
1K
Media
9
Statuses
361
Co-founder and Chief AGI Scientist, Google DeepMind
London
Joined May 2009
Learn:.1. linear algebra well (e.g. matrix math).2. calculus to an ok level (not advanced stuff).3. prob. theory and stats to a good level.4. theoretical computer science basics.5. to code well in Python and ok in C++. Then read and implement ML papers and *play* with stuff! :-).
@ShaneLegg Hey Shane I’m currently 17 from London England and am very passionate about AI, also learning about in-depth human needs. What would be the 5 pieces of advice and tips you would give to a young person like me?.
19
368
1K
I've been saying this within DeepMind for at least 10 years, with the additional clarification that it's about cognitive problems that regular people can do. By this criteria we're not there yet, but I think we might get there in the coming years.
Pragmatically, we can say that AGI is reached when it's no longer easy to come up with problems that regular people can solve (with no prior training) and that are infeasible for AI models. Right now it's still easy to come up with such problems, so we don't have AGI.
52
78
822
I remember discussing AGI risks with @DarioAmodei over dinner in New York in 2009, when we were still both students. And I was discussing the topic in online forums at least 7 years earlier than that. This is clearly not a new concern for us.
Agreed. Same for AI Risks. It's simply false that AI company leaders like @sama and @DarioAmodei talk about AI risks to make their companies/tech seem more powerful/exciting/lucrative. I am shocked how often I see people alleging that. I know both care deeply about AI risks.
17
31
578
Our research project SIMA is creating a general, natural language instructable, multi 3D game-playing AI agent. The agent can carry out a wide range of tasks in virtual worlds, making AI more adaptable, helpful & fun!
25
81
491
Wow. It seems like just yesterday (in reality more like 5 years ago) when many AGI skeptics were saying that superintelligence was not coming in the next century. How times have changed.
By "not any time soon", I mean "clearly not in the next 5 years", contrary to a number of folks in the AI industry. Yes, I'm skeptical of quantum computing, particularly when it comes to its application to AI.
24
51
457
I first met @DemisHassabis in 2009 and we bonded over a shared vision of the future, where AI helped scientists make incredible discoveries & new breakthroughs. AlphaFold has made that vision a reality. Massive congrats to Demis and John Jumper on the Nobel Prize!.
Winning the @NobelPrize is the honour of a lifetime and the realisation of a lifelong dream - it still hasn’t really sunk in yet. With AlphaFold2 we cracked the 50-year grand challenge of protein structure prediction: predicting the 3D structure of a protein purely from its.
7
27
450
#GeminiAI is another solid step towards #AGI. Huge congrats to everyone at @GoogleDeepMind who made this amazing milestone happen… and we’re just getting started :-).
17
43
423
I am so humbled to have been awarded a CBE, something I never would have imagined when I was growing up in Rotorua. This is a testament to the thoughtful, pioneering team at @DeepMindAI — the scientific breakthroughs we’ve already achieved together, and those still to come!.
29
28
383
AGI skeptics love using totally bizarre definitions of AGI to argue their case. For a sensible definition, consider the Competent AGI level from the Levels of AGI paper by @merrierm et. al. Next time you see a silly definition, point to this paper.
32
64
349
Congratulations to our team at @GoogleDeepMind who have developed #GraphCast. It’s currently the most accurate 10-day global weather forecasting system in the world, and can predict extreme weather with unparalleled speed. Read more about it here:
8
49
338
As someone who still vividly remembers trying to solve crazy hard geometry problems at the New Zealand IMO training camp in Christchurch way back in 1990. it kind of blows my mind to see how good AI has become at this! AGI keeps getting closer.
15
40
340
Who's starting to feel the AGI?.
Google just crushed ChatGPT with Gemini 2.0. Live feeds. Real-time answers. Almost zero latency. Here’s why it’s blowing everyone’s mind:. (#3 is crazy)
29
13
327
A new paper is out on the definitions & levels of AGI, written by @MerrieRM, @JaschaSD, Noah Fiedel, @TrisWarkentin, @AllanDafoe, @AleksandraFaust, @Clmt & myself:
13
72
312
Ever since the early days of my PhD, AI safety has been a personal passion of mine. I hope this resource will be one of many ways that we continue to engage with the safety-interested community as we collectively build towards safe and beneficial AI.
4
94
302
"Serious scientists" are often deeply sceptical. As a result they are often the last people to accept that something revolutionary is happening.
15
27
262
Today our team @GoogleDeepMind announced the Frontier Safety Framework — a set of exploratory protocols to proactively identify future AI capabilities, and put in place mechanisms to detect them. Great efforts from @ancadianadragan, @AllanDafoe & others
12
37
208
Many thanks to @Dwarkesh_sp for the great interview questions last week. I was definitely a bit jet-lagged, but hopefully my answers were still clear enough.
I had a lot of fun chatting with @ShaneLegg - Founder & Chief AGI Scientist, Google DeepMind. We discuss:. - Why he expects AGI around 2028.- How to align superhuman models.- What new architectures needed for AGI.- Has Deepmind sped up capabilities or safety more?.- Why
15
31
237
This Oxford/DeepMind/CIFAR paper on automatic lipreading is amazing and check out the video
6
162
228
8 years ago, millions of people around the world watched Go master Lee Sae Dol take on @GoogleDeepMind’s #AlphaGo. It was an amazing event that I’ll never forget. Very cool to now watch Lee Sae Dol reflecting on his experience ->
7
28
230
This is also my experience. The fact that AI risk critics often label such people "doomers", even when they know these people consider a positive outcome to be likely, tells you a lot about the critics.
Most people working on reducing the risk of AI catastrophe think the risk is too high but <<50%; many of them do put their money where their mouth is and are very *long* AI stocks, since they're way more confident AI will become a bigger deal than that it'll be bad.
14
16
229
Gemini Deep Research is a game changer, in my opinion.
Gemini Deep Research mode is one of the features we released, that's super useful day to day. Here I just asked for a summary of how our new Gemini Multimodal Live API got received. it crawled the web, built up a great report, with references, and I can just share it via
9
19
210
I signed this letter as I believe that AI is an exceptionally powerful technology that must be handled with great care.
We’ve released a statement on the risk of extinction from AI. Signatories include:.- Three Turing Award winners.- Authors of the standard textbooks on AI/DL/RL.- CEOs and Execs from OpenAI, Microsoft, Google, Google DeepMind, Anthropic.- Many more.
19
21
194
That our new Flash model reached such a high level of performance. I have to admit, that took me a bit by surprise. Really impressive. And the new Astra demos are mind blowing. Sci-fi is quickly becoming reality.
Welcome to the world, Gemini 2.0 ✨ our most capable AI model yet. We're first releasing an experimental version of 2.0 Flash ⚡ It has better performance, new multimodal output, @Google tool use - and paves the way for new agentic experiences. 🧵
8
15
203
There is a lot more to intelligence than just academic smarts.
If it were true that raw intelligence was sufficient for a human to want to dominate others and succeed at it, then Albert Einstein, Richard Feynman, Leonard Euler, Niels Abel, Kurt Gödel and other scientists would have been both rich and powerful, and they were neither. 1.
17
10
196
Fantastic results from the @DeepMind team at #CASP14 today. I'm really proud to see how AI is starting to help us understand some of the world’s most difficult and important scientific problems. #AlphaFold.
5
14
194
@rajiinio After more than a decade of being told that nobody serious actually believes in catastrophic AGI risks, I think it's fair game to now point out that that's not true.
5
15
179
AlphaFold Server will allow scientists to generate molecular complexes with just a few clicks! This will dramatically accelerate progress. #AlphaFold 3
6
20
179
Excited to see @DeepMindAI safety team develop this promising approach for tackling the important problem of agent alignment!.
Scalable agent alignment via reward modeling: a research direction. Paper: Blog post: @janleike @tom4everitt @MiljanMartic @v_maini @ShaneLegg
2
40
176
Huge congratulations to the #AlphaFold team! It's hard to believe how amazingly far computational protein folding has come in recent years. Wow!.
Today with @emblebi, we're launching the #AlphaFold Protein Structure Database, which offers the most complete and accurate picture of the human proteome, doubling humanity’s accumulated knowledge of high-accuracy human protein structures - for free: 1/
1
18
165
Another significant step up with our Gemini Flash models. As always, more will be coming.
Our latest update to our Gemini 2.0 Flash Thinking model (available here: scores 73.3% on AIME (math) & 74.2% on GPQA Diamond (science) benchmarks. Thanks for all your feedback, this represents super fast progress from our first release just this past
8
8
178
On the cover of @ScienceMagazine today, #AlphaMissense from @GoogleDeepMind has now categorised 89% of all missense mutations vs 0.1% prev. categorised by human experts. This AI breakthrough will advance our understanding of diseases like cancer & MND.
6
33
156
Nothing at @GoogleDeepMind could be more important than AGI Safety and I don't know of anyone better suited to lead our work in this critical area. It is with great pleasure that I welcome Anca!.
So excited and so very humbled to be stepping in to head AI Safety and Alignment at @GoogleDeepMind. Lots of work ahead, both for present-day issues and for extreme risks in anticipation of capabilities advancing.
4
12
151
While I think the Turing Test is flawed as a robust test of AGI (for various well documented reasons), I still think it's really important in another way: when agents can pass something like a Turing Test, that's the point when many people will start to intuitively "feel the AGI".
12
13
151
Huge congrats to the Gemini team @GoogleDeepMind Another step towards AGI.
Exciting News from Chatbot Arena!. @GoogleDeepMind's new Gemini 1.5 Pro (Experimental 0801) has been tested in Arena for the past week, gathering over 12K community votes. For the first time, Google Gemini has claimed the #1 spot, surpassing GPT-4o/Claude-3.5 with an impressive
13
13
146
@Simeon_Cps @GaryMarcus By AGI I mean: better than typical human performance on virtually all cognitive tasks that humans can typically do. I think we're not too far off. That some people now want to set the AGI bar so high that most humans wouldn't pass it, just shows how much progress has been made!.
12
22
145
I'm so proud of this astonishingly beautiful work! Huge congratulations to the team at DeepMind that worked on this.
The full peer-reviewed @sciencemagazine evaluation of #AlphaZero is here - a single algorithm that creatively masters chess, shogi and Go through self-play
1
28
136
The creators of the #AlphaGoMovie have made a new doc #TheThinkingGame covering the story of @GoogleDeepMind, AGI & AlphaFold. Looking forward to the premiere tonight at #Tribeca2024 and Q&A with @DemisHassabis, @DarrenAronofsky & director Greg Kohs
Come and hear @GoogleDeepMind CEO & AI pioneer @demishassabis in conversation with director @DarrenAronofsky about AI, @thinkgamefilm and the future at #Tribeca2024:
5
14
138
Today @Google @Anthropic, @Microsoft and @OpenAI are launching the Frontier Model Forum — a new industry body focused on ensuring safe & responsible development of frontier AI models. Link:
4
31
136
My recent TED interview with @TEDchris is now out! I started watching TED talks way back during my PhD, about 20 years ago. It's a bit surreal to actually be in one myself :-).
Excited about Google's Gemini? Earlier this year, Head of TED @TEDChris sat down with Google DeepMind co-founder @ShaneLegg to talk about some of his team’s groundbreaking work on AI models, artificial general intelligence (AGI) and what’s at stake:
10
11
133
Huge congrats to the organisers of the #AIAlignment Workshop at #NeurIPS2023 After being a niche community for years, it’s now like a regular academic workshop with famous professors, lots of junior professors & their students, and people in industry. And some outstanding talks!.
2
7
109
I am super proud of this amazing work by our @DeepMind_Health team working with London's @Moorfields eye hospital.
1
28
102
@XiXiDu Still seems about right to me. Of course this now means a 50% chance of AGI in the next 3 years!.
5
16
109
Huge congrats Demis. It is well deserved!.
Delighted and honoured to receive a Knighthood for services to AI. It’s been an incredible journey so far building @GoogleDeepMind over the past 15 years, helping accelerate the field and grow the UK & global AI ecosystems. Thanks to everyone who helped make this dream possible!.
1
3
103
A great way to celebrate Gemini's first birthday!.
Big news on Chatbot Arena 🔥. The new @GoogleDeepMind model gemini-exp-1206 is crushing it, and the race is heating up. Google is back in the #1 spot 🏆overall and tied with O1 for the top coding model! . Highlights (improvement since gemini-exp-1121 in parentheses). - First
5
5
104
Honoured to be included in @TIME’s #TIME100ArtificialIntelligence list, alongside my @GoogleDeepMind colleagues @DemisHassabis, @LilaIbrahim, @Pushmeet & @Shakir_za. For AGI to be positive, systems must be not only very clever, but also profoundly ethical
7
4
100
Awesome stuff. Pushing the boundaries of AI to take on serious problems like cancer is what DeepMind is all about!
Read more about our @uclh collaboration using AI to improve radiotherapy treatment for cancer patients
3
36
95
@AndrewCritchPhD I think AlphaFold created new protein structure knowledge and AlphaGo created new Go move knowledge, e.g. move 37. LLMs not so much; more blending existing knowledge. But this will soon pass and search applied to powerful foundation models will clearly produce new knowledge.
7
9
96
It was amazing to meet ~100 of the scholars @GoogleDeepMind has funded to receive AI education & hear their impact across a huge range of fields. I look forward to staying connected through our new Leader Alumni Network & continuing to empower the next generation of AI leaders!
4
12
95
MIT Tech Review article about this paper I recently co-authored on the definition and levels of AGI.
Google DeepMind wants to define what counts as artificial general intelligence
6
14
90
If you're interested in AGI Safety, check out this new collection of grid worlds (blog+paper+code). While they are as simple as we could make them, the problems they illustrate could manifest in rather subtle ways in the future.
Using simple environments to illustrate technical AI safety problems. Read the blog:.
1
30
92
The story behind AlphaGo is now available online :-)
#AlphaGo Movie is now available to rent and buy on @GooglePlay & heading to other on-demand platforms on 8 December!
0
26
85
Another step toward making AI more flexible and general
New paper: an agent that exploits interpretable vision to learn how to act robustly in novel environments
1
36
77
This is natural given the vast quantity of data suitable for SSL. Nevertheless, I'm predicting a bit of a comeback for RL as we try to shape and refine these systems for various applications.
The impact of RL has been exactly as small as I had predicted. The impact of Self-Supervised Learning has been even bigger that I had predicted.
6
11
77
And for the record, I don't think current LLMs do pass the Turing Test. They are getting much closer, obviously, but after a few minutes of asking the right kinds of questions they start making mistakes that no competent human would make.
14
6
73
Not sure where to send your new research paper on AI Safety? Check out this brand new AAAI/ACM conference:
2
33
75
We’ve come a long way since I started talking about Artificial General Intelligence (#AGI) back in 2001. Thanks @guyraz for having me on @HowIBuiltThis. Was nice to dive into my roots in AI & explore how we can better understand its value and challenges
6
10
72
I still *love* this paper's title, not to mention the great results too :-)
Code for our paper Learning to Learn by Gradient Descent by Gradient Descent is now on github.
0
23
71
AI is enabling significant scientific progress, and we’re only just getting started! I believe that AI, and AGI, are going to be deeply transformative for many fields of science in the coming decades.
In @FortuneMagazine, our CEO @DemisHassabis and @Google SVP Research, James Manyika, share why they believe AI has the potential to be a transformative tool for scientific discovery - helping us answer some of the most fundamental questions about the universe. ↓
5
7
73
When something this simple is this powerful, you're probably onto something
Excited about our new paper on relational reasoning, with @dnraposo @PeterWBattaglia and others at DeepMind.
0
15
68
@timnitGebru What an astonishing misrepresentation. These quotes were to highlight the problematic history of intelligence testing. I then point out that intelligence tests can easily be biased by the culture, race, gender or social class of the test's creator.
7
1
65
And this slightly more practical followup with Joel Veness. Foundational ideas at #DeepMind
Just reread @ShaneLegg + Hutter's classic paper on defining machine intelligence - def. recommend, lots of insights:
0
25
65
I have to admit, I've become quite attached to these little walkers :-) The obstacles they learn to overcome are really amazing!
Agents having fun in Parkour! Cool paper from colleagues at DeepMind
1
19
59
This is a truly wonderful step forward for artificial intelligence. Long live the Partnership on AI and may it stay true to its mission.
1/4 Very proud to announce the ‘Partnership on Artificial Intelligence to Benefit People and Society’ #PAI.
3
26
60
Whoo hoo, big congratulations to the teams at @DeepMind and @Blizzard_Ent who have been working so hard to make this happen!
The wait is over. Introducing SC2LE - an RL environment based on StarCraft II from DeepMind and @Blizzard_Ent
2
12
58
@GaryMarcus @Simeon_Cps 35% vs 80% in 13 years isn't soooo different — e.g. if it happened towards the end of this period then neither of us will be especially surprised. For the next 3 years I'm at about 30%. I.e. I'd be slightly surprised, while you'd be totally shocked :-).
2
8
52
Good BBC article about some of the Technical AI Safety research that we're doing in collaboration with FHI
5
35
52
It's been a great year with an amazing team! I'm already looking forward to 2018.
2
9
52
Nice overview of our AI Safety Gridworlds paper on Computerphile
1
21
46
Looking through all the ICLR submissions, what was a blizzard of new deep learning results has now turned into an avalanche!.
1
15
48
I enjoyed speaking to @strwbilly about AGI for this story.
If we had machines that could think like us or better, then maybe we’d stand a better chance at solving the world’s biggest problems. That’s been the guiding vision of AI research since the earliest days—and remains its most divisive idea.
2
8
45
@daniel_271828 @BogdanIonutCir2 @XiXiDu I think this is currently one of the most important questions. Investigating this is one of my top priorities.
1
2
47
Indeed, this was something that I had in mind when I suggested the term "AGI" to @bengoertzel for his book, back in 2001. But mostly, we just wanted to see more research on AI systems that aim to be quite general, for example, as general as human intelligence.
Also worth noting: "strong AI" was a term coined by John Searle to refer to belief that “the appropriately programmed digital computer does not just simulate having a mind; it literally has a mind.” Thus "Strong AI" is not a synonym for "AGI", but rather a philosophical stance.
3
7
44
It's really great that we can play a part in training the next generation of machine learning experts! #UCL #DeepMind
UCL students can now learn state-of-the-art #ArtificialIntelligence with our @DeepMindAI partnership #AI
1
13
43
#GNoME is a new AI tool from @GoogleDeepMind that has helped discover 2.2 million new crystals — I'm super excited to see what interesting properties some of these new materials will have!
1
4
43
@ESYudkowsky @Simeon_Cps Roughly exponential growth in compute and available data causing the margin value of more scalable and capable algorithms to become extremely high, driving innovation hard. During 2020's should then be possible to train huge models on far more data than a human could experience.
1
1
40
Thanks to @businessinsider for recognising our work and putting together a fun evening — I always wanted to see the inside of The Gherkin!
1
9
40
It was a real pleasure to present the #MaxwellLecture this evening at King's College London. James Clerk Maxwell is one of my scientific heroes and he did some of his most brilliant work right here at King's!
We have a full house for the #MaxwellLecture with @ShaneLegg from @DeepMindAI organised by @KCLPhysicsSoc
1
7
39
Cold, grey & damp, avocado toast, smokey bacon & a flat white, classic punk, rap & blue grass playing in the café: a North London state of mind.
3
35
36
Great conversation with @ReedAlbergotti about #AGI at @sxsw The possibilities opening up before us are immense.
3
5
40
I really enjoyed this film and thought director Greg Kohs did an amazing job. Hope you all get to see it when it's released more widely!
1
9
38
They said this wouldn't happen for a long time. but it just did! Well done team and thanks to Sedol for fighting hard to the end #AlphaGo.
2
14
37
A great intro to our mission here at #DeepMind and the important problems we want to use technology to tackle!
Here’s my talk announcing our exciting research effort on diabetic retinopathy with @Moorfields .
0
10
35
Ten years later and I still like that cover design :-).
3
3
36
Had a great time at Women in Machine Learning today and met some really interesting researchers. Thanks to the organisers for putting on this awesome event. Well done #WiML2017
2
8
35
@TwoMinutePapers once again manages to pull the key ideas out of a paper in an entertaining way. This time about SIMA, our new language instructable multi-3D-game playing AI. As the review notes, the real significant of this lies in what will come next.
5
6
35
I've had a fantastic week in Aotearoa! It was a huge honour to receive a Distinguished Alumni Award @Waikato and such a pleasure to see so many familiar faces at my talk :-).
3
1
35
I had a great conversation with @ShirinGhaffary yesterday, all about the definition of AGI.
More tech companies are trying to develop artificial general intelligence. There’s just one problem: no one can quite agree on what it is
3
4
34
Big thanks to everyone from @UCLEntrepreneur for organising my lecture about DeepMind's research this evening. It was great meeting you all!.
2
7
33
I can't wait to see how AI research progresses on StarCraft2 in the coming years. It's a seriously hard challenge :-)
StarCraft keeps coming back in my life : ) Together with Blizzard, we are opening StarCraft2 for AI research! @Blizzard_Ent @DeepMindAI
1
8
33
It's been another amazing week of Go playing :-) Looking forward to the coming official matches!
0
12
31
@NandoDF @ylecun @AndrewYNg Change of mind? Earlier this year you signed the CAIS letter which said: Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.
2
1
30
@NandoDF @ylecun @AndrewYNg "belief endorsed by a few" seems important to me. I say this because in my experience most people concerned with catastrophic risks from future AGI systems also believe that present day problems and risks are also very important — a point which is often lost in Twitter debates.
1
2
30
@NandoDF @ylecun @AndrewYNg So you think extinction from AI is “cultist AI long term speculation” and you also think it “should be a global priority”? That's not a perspective that I often come across.
4
0
29
@ESYudkowsky @Simeon_Cps I also thought neuroscience would provide more AGI design hints. It hasn't much, other than at a very high level, but I now suspect that the lack of this driver of progress isn't going to matter much.
2
2
28