Helen Toner @hlntnr Twitter profile | Pikagi

Pikagi

Helen Toner

@hlntnr

21,595

Followers

1,203

Following

235

Media

1,508

Statuses

Interests: China+ML, natsec+tech, brains+words+absurdity | Current: @CSETGeorgetown (opinions my own) | Former: @open_phil

https://t.co/iioh86SDWB

Joined November 2014

Don't wanna be here? Send us removal request.

Pinned Tweet

@hlntnr

Helen Toner

2 months

Seems like everyone wants to know more about how large language models work, but existing explainers often: -Are too shallow or too technical -Focus on how LLMs predict the next word, which is only part of the story So we wrote 3 explainers of our own! 🧵

Tweet card media

The Surprising Power of Next Word Prediction: Large Language Models Explained, Part 1 | Center for...

Large language models (LLMs), the technology that powers generative artificial intelligence (AI) products like ChatGPT or Google Gemini, are often thought of as chatbots that predict the next word....

cset.georgetown.edu

8

62

284

Last Seen Profiles

@latkedelrey

@looloo2ll

@trillxsashaa

@remiwolf

@gbpkay

@ItsWebbJP

@julissa_eu

@flo_wershower

@hocca_official

@lecmonza

@gonzalezpons

@druzinska

@mywife_latex

@jamesocromwell

@ecto_fun

@FantomApeTalks

@HowardU

@hakuoh_soldiers

@kt_fps

@BrawlStars

@Blood_NSFW

@princessemathil

@tekno_71

@MrLePresiden

@KarmineCorp

@MiRXESInsights

@DissidentSoaps

@timetojiyong

@raqueebalam

@SaeedDiCaprio

@umzemaneai

@VCITW

@V_USA

@ZiFMStereo

@f0rever4th

@JimLoftus4

@hlntnr

Helen Toner

5 months

Today, I officially resigned from the OpenAI board. Thank you to the many friends, colleagues, and supporters who have said publicly & privately that they know our decisions have always been driven by our commitment to OpenAI’s mission. 1/5

865

606

7K

@hlntnr

Helen Toner

6 months

And now, we all get some sleep😴

@OpenAI

OpenAI

6 months

We have reached an agreement in principle for Sam Altman to return to OpenAI as CEO with a new initial board of Bret Taylor (Chair), Larry Summers, and Adam D'Angelo. We are collaborating to figure out the details. Thank you so much for your patience through this.

6K

13K

67K

568

165

4K

@hlntnr

Helen Toner

1 year

If you spend much time on AI twitter, you might have seen this tentacle monster hanging around. But what is it, and what does it have to do with ChatGPT? It's kind of a long story. But it's worth it! It even ends with cake 🍰 THREAD:

Tweet media one

42

591

3K

@hlntnr

Helen Toner

5 months

I have enormous respect for the OpenAI team, and wish them and the incoming board of Adam, Bret and Larry all the best. I’ll be continuing my work focused on AI policy, safety, and security, so I know our paths will cross many times in the coming years. 5/5

108

32

1K

@hlntnr

Helen Toner

5 months

To be clear: our decision was about the board's ability to effectively supervise the company, which was our role and responsibility. Though there has been speculation, we were not motivated by a desire to slow down OpenAI’s work. 3/5

38

68

883

@hlntnr

Helen Toner

7 months

My new obsession is asking DALL-E 3 for diagrams of complicated things. This is "please create a clearly labeled diagram of how reverse genetics works"

Tweet media one

Tweet media two

Tweet media three

Tweet media four

25

56

752

@hlntnr

Helen Toner

5 months

When I joined OpenAI’s board in 2021, it was already clear to me and many around me that this was a special organization that would do big things. It has been an enormous honor to be part of the organization as the rest of the world has realized the same thing. 4/5

17

17

629

@hlntnr

Helen Toner

5 months

Much has been written about the last week or two; much more will surely be said. For now, the incoming board has announced it will supervise a full independent review to determine the best next steps. 2/5

9

19

627

@hlntnr

Helen Toner

6 months

Thread of my quick reactions to the AI executive order: (the full text doesn't appear to be out yet, so this is based on the factsheet for now - but I may add more thoughts later once the full thing is up)

Tweet card media

FACT SHEET: President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial...

Today, President Biden is issuing a landmark Executive Order to ensure that America leads the way in seizing the promise and managing the risks of artificial intelligence (AI). The Executive Order...

www.whitehouse.gov

18

127

482

@hlntnr

Helen Toner

4 years

In less dire news- did you know that when you finish your PhD at some German universities, your lab decorates a personalized hat & wagon for you and your professor tows you around the campus? Relatedly: I'm bursting with pride for my brilliant partner, who defended yesterday! 🎉

Tweet media one

Tweet media two

5

51

458

@hlntnr

Helen Toner

11 months

New piece from @jennywxiao , @jjding99 and me pushing back on the claim that we can't regulate AI because that would just let China pull ahead. This is not a good argument! 4 reasons why, though first a caveat: 🧵

@ForeignAffairs

Foreign Affairs

@ForeignAffairs

11 months

“Even if it were true that new AI regulations would slow innovation in the United States—and it very well may not be—China does not appear poised to surge ahead.”

5

32

97

49

131

418

@hlntnr

Helen Toner

1 year

Calling for a 6-hour pause on AI discourse so that we can all take a deep breath, have a cup of tea, go on a walk to the park

10

42

383

@hlntnr

Helen Toner

3 years

Frustrated by coverage of the vaccine rollout that makes it sound like running out of vaccine is purely a bad thing. Hospitals should be doing everything in their power to run out of vaccine right now! Good job, hospitals with empty vaccine lockers! 🏆 1/3

Tweet media one

5

39

353

@hlntnr

Helen Toner

3 years

After more than 5 years of long distance, today my partner arrived in DC and in January he'll start a postdoc at Georgetown. Still pinching myself, but it appears to actually be real.

Tweet media one

14

0

333

@hlntnr

Helen Toner

2 months

A statement from Helen Toner and Tasha McCauley:

Tweet media one

37

44

324

@hlntnr

Helen Toner

3 years

No big deal, just sitting here alone in my apartment sobbing because I got an email from USCIS that I *think* means my green card has been approved 😭

20

0

318

@hlntnr

Helen Toner

3 years

✨Personal news alert✨ On Tuesday my partner of ~7 years and I got married in a teensy ceremony on Georgetown campus. We told very few people in advance and are looking forward to celebrating with friends when that's a thing again, but for now we're thrilled to bits. 🥰

Tweet media one

Tweet media two

Tweet media three

Tweet media four

36

1

307

@hlntnr

Helen Toner

8 months

Please, I'm begging you, stop talking about "short-term and long-term risks" from AI. It's wrong on both ends - 1) "Short-term" "risks" is a dreadful way to describe harms that are already happening and might continue long into the future...

2

52

277

@hlntnr

Helen Toner

2 years

Personal update: made a human! She's the greatest and the cutest. I'll probably be tweeting (even) more intermittently than normal for the next few months; try not to kill each other while I'm gone 💕

Tweet media one

27

2

271

@hlntnr

Helen Toner

10 days

I spoke at the main TED conference in Vancouver a couple of weeks ago! Was a surreal but fun experience. My talk is up today—it's about how to govern AI, even though we struggle to understand and predict it. Link below, 🧵 on the main points:

Tweet card media

How to govern AI — even if it's hard to predict

No one truly understands AI, not even experts, says Helen Toner, an AI policy researcher and former board member of OpenAI. But that doesn't mean we can't govern it. She shows how we can make smart...

17

60

280

@hlntnr

Helen Toner

3 years

I'm thrilled and incredibly honored to share this news! I can't wait to work with @sama , @gdb , @ilyasut and the rest of the team to contribute to the mission of building AI that benefits all of humanity.

@OpenAI

OpenAI

3 years

We're excited to announce @hlntnr of @CSETGeorgetown is joining our board of directors. Her deep understanding of AI policy will help us achieve our mission to deploy safe and responsible general-purpose AI.

83

46

346

20

14

255

@hlntnr

Helen Toner

1 month

If accurate, this is appalling. Describes how AI-based targeting lets targeting happen faster, which creates pressure to move faster, which means humans just rubber-stamp the AI suggestions even though they're known to have a 10% error rate.

Tweet card media

‘Lavender’: The AI machine directing Israel’s bombing spree in Gaza

The Israeli army has marked tens of thousands of Gazans as suspects for assassination, using an AI targeting system with little human oversight and a permissive policy for casualties, +972 and Local...

13

55

247

@hlntnr

Helen Toner

1 year

Back in 2016, Yann LeCun (Chief AI Scientist at FB) presented this slide at NeurIPS, one of the biggest AI research conferences. Back then, there was a lot of excitement about RL as the key to intelligence, so LeCun was making a totally different point...

Tweet media one

2

16

237

@hlntnr

Helen Toner

5 years

Great/scary site: If you've already played around with AI-generated images a bunch, try doing this without looking closely at the backgrounds (i.e. self-imposed hard mode 🧐)

20

100

218

@hlntnr

Helen Toner

1 year

...Maybe the bulk of what's going on is an inhuman Lovecraftian process that's totally alien to how we think about the world, even if it can present a nice face. (Note that it's not about the tentacle monster being evil or conscious—just that it could be very, very weird.)

Tweet media one

6

14

214

@hlntnr

Helen Toner

2 years

One of my favorite quotes on AI is from Dijkstra: "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." But not just as a clever comeback! 1/3

@random_walker

Arvind Narayanan

2 years

Debating whether today's AI tools are truly intelligent is about as useful as debating whether a drone is a bird. If you define a bird to be anything of a certain size that flies, then sure. But it's also obvious that drones can't do most of the other things that birds can do.

9

65

345

5

20

199

@hlntnr

Helen Toner

1 year

Now we know what all the words mean, the picture should make more sense. The idea is that even if we can build tools (like ChatGPT) that look helpful and friendly on the surface, that doesn't mean the system as a whole is like that. Instead...

Tweet media one

4

17

206

@hlntnr

Helen Toner

7 months

um ah "please create a detailed diagram of how to determine a horse's age according to its teeth"

Tweet media one

Tweet media two

Tweet media three

Tweet media four

11

15

202

@hlntnr

Helen Toner

4 years

I unironically love that this @ESYudkowsky quote has become a BLM slogan. Also spotted in my local park recently:

Tweet media one

@MemphisQuinn

You know, I’m something of a Drapetomaniac myself

4 years

Yup

Tweet media one

106

51K

128K

5

17

196

@hlntnr

Helen Toner

1 year

Can't escape the feeling that people are living in two different realities re: where AI is heading over the next 5-10 years. All I can say is, if you feel like you know what to expect then now is the time to get predictions on the record - what will & won't AI be able to do?

45

8

191

@hlntnr

Helen Toner

8 months

2) And the people most worried about "long-term" risks are worried bc they think they might happen soon! Not ideal. If you really want to talk about 2 categories, I suggest "existing harms" and "anticipated risks" (or "speculative risks"). Thank you for coming to my TED talk.

5

18

181

@hlntnr

Helen Toner

1 year

First, some basics of how language models like ChatGPT work: Basically, the way you train a language model is by giving it insane quantities of text data and asking it over and over to predict what word[1] comes next after a given passage. Eventually, it gets very good at this.

1

3

155

@hlntnr

Helen Toner

2 years

Characteristically wise words from @KelseyTuoc on psychologically coping over the next couple of months:

Tweet media one

Tweet media two

5

23

156

@hlntnr

Helen Toner

1 year

...Namely, that RL was only the "cherry on top," whereas unsupervised learning was the bulk of how intelligence works. To an AI researcher, the labels on the tentacle monster immediately recall this cake, driven home by the cheery "cherry on top :)" END THREAD

1

1

148

@hlntnr

Helen Toner

1 year

And then you do reinforcement learning on *that*, so the AI is trying to maximize how much the humans will like text it generates, based on what it learned about what humans like. So that's ✨RLHF✨ And now we can go back to the tentacle monster!

1

2

147

@hlntnr

Helen Toner

1 month

This is some of the darkest predictions of AI in warfare, coming true in 2024.

Tweet media one

8

30

144

@hlntnr

Helen Toner

1 year

How it works, very roughly, is that you give the model some prompts, let it generate a few possible completions, then ask a human to rank how good the different completions are. Then, you get your language model to try to learn how to predict the human's rankings...

3

4

140

@hlntnr

Helen Toner

4 months

V interesting devt from China: 1st CCP-approved dataset for training large language models⬇️ Was predictable they'd do this, but I have so many qs 👀 -Depending how you count, this seems to be ~100-1000x too small to train a cutting edge model. Is more coming? How much more? 1/4

@mattsheehan88

Matt Sheehan

5 months

🇨🇳starting to offer officially-endorsed data sets for training LLMs. Industry assocation under CAC offering one that's: 120 gbs 100m data points (数据条) 50b tokens (llama 13B = 1T tokens) Lowers barrier for building smaller, politically safe LLMs

1

8

45

5

23

140

@hlntnr

Helen Toner

1 year

Another trick is called "reinforcement learning from human feedback," or RLHF. The way reinforcement learning *usually* works is that you tell an AI model to maximize some kind of score—like points in a video game—then let it figure out how to do that. RLHF is a bit trickier:

1

6

140

@hlntnr

Helen Toner

1 year

But it turns out models trained that way, by themselves, aren't all that useful. They can do some cool stuff, like generating a news article to match a lede. But they often find ways to generate plausible-seeming text completions that really weren't what you were going for.[3]

Tweet media one

Tweet media two

1

2

139

@hlntnr

Helen Toner

1 year

Unlike the original training, this approach is "supervised" because the data you're using is structured as well-labeled input-output pairs. So you could also call it ✨supervised fine-tuning✨

1

4

138

@hlntnr

Helen Toner

5 years

Thrilled to finally be able to share the founding of CSET - the Center for Security and Emerging Technology. I could not be more excited to work with Jason and the incredible team we've started building. 1/3

Tweet card media

Georgetown launches think tank on security and emerging technology

The center is backed by a $55 million grant from an organization funded by one of Facebook’s co-founders.

www.washingtonpost.com

3

39

138

@hlntnr

Helen Toner

1 year

...and tree poem example from [4] The tentacle monster I used here is from ; I think the original idea is from Thanks for reading!

@TetraspaceWest

Tetraspace 💎🔎

@TetraspaceWest

1 year

Tweet media one

16

72

865

11

3

132

@hlntnr

Helen Toner

1 year

This training is a type of ✨unsupervised learning✨[2] It's called that because the data (mountains of text scraped from the internet/books/etc) is just raw information—it hasn't been structured and labeled into nice input-output pairs (like, say, a database of images+labels).

1

3

135

@hlntnr

Helen Toner

1 year

So researchers figured out some ways to make them work better. One basic trick is "fine-tuning": You partially retrain the model using data specifically for the task you care about.

2

4

133

@hlntnr

Helen Toner

24 days

This is fantastic news for NIST 🎉 I'm biased because he's a friend, but Paul Christiano is both technically brilliant and holistically super thoughtful - not to mention a pioneer in exactly the kind of frontier testing work he'll be leading at AISI.

Tweet card media

U.S. Commerce Secretary Gina Raimondo Announces Expansion of U.S. AI Safety Institute Leadership...

U.S. Secretary of Commerce Gina Raimondo announced today additional members of the executive leadership team of the U.S. AI Safety Institute (AISI), which is housed at the National Institute of...

www.commerce.gov

3

8

133

@hlntnr

Helen Toner

2 years

I will never not retweet someone breaking down an important disagreement into specific, falsifiable forecasts (even offering to bet!) 🙌👀

Tweet media one

@GaryMarcus

Gary Marcus

2 years

are you in, @elonmusk ? I say no way we will get to AGI by 2029, and lay down 5 suggestions for how we might tell:

34

31

168

7

24

127

@hlntnr

Helen Toner

6 years

Visited Xining last week, which seemed to enjoy flaunting its surveillance capabilities. Pic 1: list of cars caught parking in the bus lane. Pic 2: airport screen telling me that if I stand still long enough for the facial recog to work, it'll show me the status of my flight.

Tweet media one

Tweet media two

1

68

129

@hlntnr

Helen Toner

1 year

I'm working on an piece about how we desperately need to be able to talk about progress in AI in richer terms than "this is basically AGI" vs "this is nothing like AGI." This👇is a fantastic example of what we need more of - very worth reading.

@sleepinyourhat

Sam Bowman

@sleepinyourhat

1 year

I’m sharing a draft of a slightly-opinionated survey paper I’ve been working on for the last couple of months. It's meant for a broad audience—not just LLM researchers. (🧵)

Tweet media one

22

272

1K

2

9

118

@hlntnr

Helen Toner

1 year

If you're training a customer service bot, for instance, then maybe you pay some human customer service agents to look at real customer questions and write examples of good responses. Then you use that nice clean dataset of question-response pairs to tweak the model.

1

2

107

@hlntnr

Helen Toner

1 year

Last week, China released a draft of regulations for generative AI - big news! There's some pretty ambitious stuff in there. If you want to know more, see the analysis from me and seven others for DigiChina here 👇 Or if clicking away from Twitter is too big an ask, have a 🧵:

@DigiChn

Stanford DigiChina Project

1 year

NEW: 8 specialists on the implications and questions around China's draft rules on generative AI—from input data to accurate outputs. A Forum w/ @hlntnr @zac_haluza Yan Luo+Xuezi Dan @CovingtonLLP @mattsheehan88 @seatonyhuang Kimball Chen @China_Digital

2

20

33

4

33

107

@hlntnr

Helen Toner

5 years

A short thread on this very neat result from OpenAI. The rubik's cube is a cute touch, but the really interesting part here is: 1) They trained their models *entirely* in simulation—i.e. without needing to practice on the real robot at all.

@OpenAI

OpenAI

5 years

We've trained an AI system to solve the Rubik's Cube with a human-like robot hand. This is an unprecedented level of dexterity for a robot, and is hard even for humans to do. The system trains in an imperfect simulation and quickly adapts to reality:

247

4K

11K

2

47

103

@hlntnr

Helen Toner

1 year

A few post-thread "well, technically"s and sources: [1] Technically LMs predict what "token" comes next, but a token is usually roughly a word, so, close enough [2] Technically people often say "self-supervised learning" rather than "unsupervised learning" for LM pre-training..

1

2

103

@hlntnr

Helen Toner

3 years

wait wait wait, did we just lose the ability to use the whole concept of "meta" without it sounding like branding??

4

7

99

@hlntnr

Helen Toner

11 months

4: This one's a big one, say it with me: China👏is👏already👏regulating👏AI! People trying to be smart by saying that regulations would put the US behind are just showing how little they know about China. We're already behind them, but not on tech development—on regulating!

3

16

98

@hlntnr

Helen Toner

6 months

2) NIST is tasked with developing "standards, tools, and tests" for safe/secure/trustworthy AI. This is great, I love it, and also, where will they get the money and the people? (This would be a great place for a functioning Congress to jump in & back this up with $)

3

11

96

@hlntnr

Helen Toner

3 years

Jason has moved on from CSET to coordinate the White House's work on technology & national security! Selfishly I might be a *tiny* bit sad to lose him as a boss, but this is fantastic news for the country and the world 🚀🌏

@CSETGeorgetown

CSET

@CSETGeorgetown

3 years

CSET congratulates Founding Director Jason Matheny on his new roles in the Biden administration: Deputy Assistant to the President for Technology and National Security, Deputy Director for National Security at @WHOSTP and Coordinator for Technology and National Security at @WHNSC

8

35

223

3

3

95

@hlntnr

Helen Toner

1 year

"But wait," I hear you say, "You promised cake!" You're right, I did. And here's why—because the tentacle monster is *also* a play on a very famous slide by a very famous researcher.

1

1

93

@hlntnr

Helen Toner

2 years

Professional news: excited to share that I'm taking on a new role @CSETGeorgetown ! I'll be leading a new program to fund technical research outside CSET on questions that are important for our mission, but more technical than our usual policy-focused analysis. 1/3

1

2

92

@hlntnr

Helen Toner

2 months

5 years! It's been unbelievable to see how CSET's team and reputation has grown. To celebrate, here are 5 papers/products, 1 from each year of CSET's existence, that I love (and that exemplify the work we do).

@CSETGeorgetown

CSET

@CSETGeorgetown

2 months

🎉🎈Happy 5th Birthday to Us! 🎈🎉 In 5 short years, we’ve: ✍️Written 250+ reports 📊Published 100+ data visualizations 🌐Translated 500+ documents And so much more! Our new annual report runs through all the biggest highlights from 2023 and beyond.

0

6

24

4

6

92

@hlntnr

Helen Toner

6 years

Demography facts according to this gif, from most to least obvious: 1) the 1-child policy happened 2) the great leap forward sure also happened 3) the results... somehow strongly resemble each country's traditional architecture? 🧐

@simongerman600

Simon Kuestenmacher

@simongerman600

7 years

Comparing population pyramids of #China and #India reveals different challenges for the two emerging giants.

23

484

606

3

18

90

@hlntnr

Helen Toner

5 years

If you've been thinking "this Center for Security and Emerging Technology place sounds cool, when are we going to get to see what they've been working on?" then boy, September is the month for you. 3 reports coming next week, plus 1st edition of our newsletter, then...

1

14

87

@hlntnr

Helen Toner

2 years

Can ML systems "think"? Are they "intelligent"? I honestly don't care. What matters is what capabilities & weaknesses they have, how we use them in practice, and what harms/benefits result. Definitional debates about what "real" intelligence is rarely help answer those Qs. 3/3

3

4

82

@hlntnr

Helen Toner

3 years

This is a race between (very roughly) 3 groups: 1) the federal government, 2) state governments, and the 3) providers actually giving vaccinations. Each group should be racing as fast as it can *not to be the bottleneck*. Hospitals running out of vaccine = group 3 is winning. 2/3

1

18

82

@hlntnr

Helen Toner

5 years

The subhead of the Thiel op-ed encapsulates everything the national security world gets wrong about AI. At its core, AI is *not* a military technology - that's *precisely* what makes it hard to make policy for. To say nothing of the many other issues with the op-ed, e.g...

Tweet media one

4

15

80

@hlntnr

Helen Toner

3 years

Exactly 2.5 years ago we huddled nervously around a table, 3 days out from launching the CSET website and announcing ourselves to the world. I can't express how lovely it is to see our amazing team—*and* smart external folks who read our work—now celebrating this milestone. 🥰

@CSETGeorgetown

CSET

@CSETGeorgetown

3 years

📣 News Release! 📣 CSET is happy to announce an increase in grant funding, bringing our total funds to more than $100 million, to continue our work at the intersection of emerging tech and security.

1

19

106

1

0

81

@hlntnr

Helen Toner

5 years

Thread of musings on sth I noticed recently: conversations about how we might find meaning in a post-work world heavily feature music and art... but I can't remember sports being mentioned even once. How come, when it provides so much meaning/community/joy to so many people? 1/5

6

19

78

@hlntnr

Helen Toner

6 months

1) Glad to see the very first point is exactly what we rec'd in a recent blog post: requiring companies to share info about so-called "frontier AI" systems. Curious how they're defining which models count. More context in our post from last week:

Tweet media one

4

8

77

@hlntnr

Helen Toner

6 months

3) Tackles fears of AI-driven biological risks by focusing on the biological side (shoring up DNA synthesis screening) - very sensible. 4) Exploring watermarks, authentication for official content, using AI to fix cyber vulns, I love it

2

3

77

@hlntnr

Helen Toner

10 days

@BasedBeffJezos @MelindaBChu1 @tsarnick Sure - likewise, would be interested to know more about your takes. I'll DM you.

8

0

79

@hlntnr

Helen Toner

4 years

Honored to be involved in creating this report, and excited to share it today! If you've ever been all 🤨 about AI ethics principles, this is for you - it's all about ways we could make verifiable claims about the trustworthiness of AI systems. 1/

Toward Trustworthy AI: Mechanisms for Supporting Verifiable Claims

The increasingly widespread application of AI research has brought growing awareness of the risks posed by AI systems, and recognition that existing regulations and norms in industry and academia are...

www.towardtrustworthyai.com

2

22

76

@hlntnr

Helen Toner

3 years

When you say "AI" in DC, most people think "big data." It's not totally wrong, but this idea lags behind progress in the field. In @sciam today, @HsjChahal and I introduce one up and coming "small data" paradigm, transfer learning: 1/4

Tweet card media

'Small Data' Are Also Crucial for Machine Learning

The most promising AI approach you’ve never heard of doesn’t need to go big

www.scientificamerican.com

2

23

75

@hlntnr

Helen Toner

11 months

3: The current hottest area of AI research—large language models—really isn't a good fit for China's hypersensitive political environment under Xi. I love the way this CEO described the problem (exaggerating for comic effect):

Tweet media one

1

9

75

@hlntnr

Helen Toner

1 year

I wrote for @TIME about an analogy to help some of the weird things about AI chatbots make more sense: Try thinking of them as "improv machines." I think this helps (more than "autocomplete" or "stochastic parrots") with a few different kinds of mistakes I see people making:

@TIME

TIME

1 year

"Try thinking of chatbots as 'improv machines.'" 📝 Helen Toner ( @hlntnr )

6

9

27

3

13

70

@hlntnr

Helen Toner

1 year

...because you could think of "chunk of text"+"next word" as an input-output pair that is automatically present in the dataset—but I guess the tentacle monster artist went with "unsupervised" to match the cake. [3] Unicorn story from ...

2

2

68

@hlntnr

Helen Toner

4 years

This extremely misleading summary of a fairly misleading WHO statement will get people killed. It's very clear that PRE-symptomatic spread is a major driver of infection, so for goodness' sake don't tell people that if they don't have symptoms they don't need to be careful.

@CNBC

CNBC

4 years

Coronavirus patients without symptoms aren’t driving the spread of the virus, the WHO says.

154

343

583

1

19

69

@hlntnr

Helen Toner

2 years

Apparently you can get one of these without filing a single USCIS form or going to an interview or even like giving fingerprints? just by being a baby?? what a scam

Tweet media one

5

0

66

@hlntnr

Helen Toner

3 years

The more shots we get in arms, the sooner we get the pandemic under control. Headlines of "hospitals are running out, oh no" lead the reader to conclude that hospitals should be retaining doses, rather than pointing at the actual bottleneck: federal supply & state distro. 3/3

0

6

62

@hlntnr

Helen Toner

3 years

Perfect demonstration of what anyone with half a clue has been saying all along: treating Chinese internet co's as uniquely irresponsible with data—and US co's as inherently trustworthy—is idiotic. The way to do better is not that complicated... (1/2)

@MaraHvistendahl

Mara Hvistendahl

@MaraHvistendahl

3 years

SCOOP: The Oracle TikTok deal was supposed to prevent TikTok from passing data to Chinese police. Turns out Oracle has been marketing its own data analytics software to...Chinese police. a.k.a. the weirdest China tech story I’ve ever written 1/1

46

1K

2K

1

17

58

@hlntnr

Helen Toner

4 years

Spot on commentary from @jackclarkSF in his newsletter, on the IB's decision to grade students with ML. 🔥: "This is a terrible idea. It's almost a made-to-order example of how you shouldn't deploy an AI system."

Tweet media one

1

17

61

@hlntnr

Helen Toner

6 years

What worries me most about this: smart young Chinese have often heard claims that Chinese censorship is harsh, but it's hard for them to know how it actually compares to other countries. Google - world's biggest, most popular internet co - being banned is clear & undeniable 1/

@rj_gallagher

Ryan Gallagher

6 years

Revealed: Google is planning to launch a censored search engine in China that will blacklist websites & search terms about human rights, democracy, political opposition, academic studies, religion, & peaceful protest:

245

2K

2K

4

27

60

@hlntnr

Helen Toner

3 years

Update 3: green card was sent to the wrong address, seemed to be lost applied for it to be reissued (takes ~12mo) went through big hassle to get a temp passport stamp in the meantime (for travel) turns out it had been returned to USCIS asked them to resend, got it yesterday 😭

Tweet media one

@hlntnr

Helen Toner

3 years

Update 1: this was a false alarm 😑 unclear why I got that email, but I still had to do a medical exam so in retrospect it couldn't have been an approval Update 2: Submitted medical a few weeks ago, then today I received the exact same email... so maybe this time for real?! 🤞🤞

5

0

42

5

0

61

@hlntnr

Helen Toner

1 year

My new response to anyone expressing skepticism about anything I say (from @ezraklein )

Tweet media one

1

2

55

@hlntnr

Helen Toner

3 years

To everyone watching the NY mayoral race, I just want to say - I know nothing about the NY board of elections, but I've watched decades of Australian ranked choice races and I PROMISE they don't have to be this much of a shitshow. 📊 Don't blame it all on ranked choice!

5

6

59

@hlntnr

Helen Toner

11 months

It would be a huge mistake to let fear of China outstripping us prevent us from regulating AI. I hope the piece can help push back against that narrative. Read it in full here (no paywall for the next few days):

Tweet card media

The Illusion of China’s AI Prowess

Regulating AI will not set America back in the technology race.

www.foreignaffairs.com

1

12

59

@hlntnr

Helen Toner

2 years

The point isn't "what a dumb comparison, swimming is so much more than what submarines do." The point is that our everyday words often bunch together multiple different concepts, and when we encounter something new (like submarines or AI), we need to disentangle & unpack. 2/3

1

3

54

@hlntnr

Helen Toner

2 years

My most mundane life hack: just clean your glasses in the sink with warm water + soap + your fingers. You don't need a fancy cloth or spray, and this will get them *way* cleaner than rubbing on your shirt. And no, they won't scratch - Big Glasses has lied to you. Try it! 🤓

@SpencrGreenberg

Spencer Greenberg 🔍

@SpencrGreenberg

2 years

A question for you: what's one of your favorite life hacks that is actually useful?

70

9

91

5

2

55

@hlntnr

Helen Toner

5 years

New article by @lorandlaskai and me on the foundations underlying China's AI push, published in the DigiChina report released yesterday: tl;dr: If chips, basic R&D, and ML frameworks are included in its ambitions to lead the world, China has a ways to go.

Tweet card media

Special Report: AI Policy and China

www.newamerica.org

2

17

55

@hlntnr

Helen Toner

3 years

My first pubs with CSET came out this week! 3 short papers intro'ing key ideas related to the sometimes-mysterious phrase "AI safety." First up is an overview, breaking down "safety" into three more specific areas: robustness, assurance, specification.

Key Concepts in AI Safety: An Overview | Center for Security and Emerging Technology

This paper is the first installment in a series on “AI safety,” an area of machine learning research that aims to identify causes of unintended behavior in machine learning systems and develop tools...

cset.georgetown.edu

1

6

54

@hlntnr

Helen Toner

3 years

today, I went into a bookstore and just browsed 💉😍💉

1

0

53

@hlntnr

Helen Toner

4 years

We honestly didn't plan this timing, but - the most recent talent report from @CSETGeorgetown is out today, comparing work/study/entrepreneur visas in the US, Canada, France, UK, and Australia: Well done @tinahuang__ for getting this out!

Tweet media one

Tweet media two

Tweet media three

2

19

52

@hlntnr

Helen Toner

4 years

Also: it's called your Doktorwagen, of course

1

0

51

@hlntnr

Helen Toner

6 months

12) Seeing a few people asking about how this compares with the EU's AI Act - short answer is, it doesn't. This isn't a law! The White House is *extremely* constrained in what it can do on its own (i.e. without Congress passing new laws)...

2

4

51

@hlntnr

Helen Toner

6 months

5) A whole section on privacy w/ big focus on privacy enhancing technologies (PETs) - awesome, I love it, my program funds stuff related to this... and also it's no substitute for a federal privacy law. But not much the WH can do about that. More on PETs:

Tweet card media

Privacy Is Power

Global digital society's operating system is overdue for an upgrade.

www.foreignaffairs.com

2

5

50

@hlntnr

Helen Toner

4 years

Update: one anxious wait is over! My niece Lily Grace was born a couple of hours ago, and is absolutely beautiful 🥰 Will see if I can get permission to share photos here.

@hlntnr

Helen Toner

4 years

I know it's not a competition for who can do the most waiting the most anxiously, BUT at midnight last night my sister went into hospital to have her first baby and we haven't heard anything since, so

2

0

39

4

0

51

@hlntnr

Helen Toner

2 years

...ok and bonus, here's what #dalle2 had to say about the situation 🚀 ("helen setting out on a thrilling new professional adventure to save the world, watercolor")

Tweet media one

0

1

52

@hlntnr

Helen Toner

6 months

-Kind of funny to see USG getting caught up in the competition of superlatives in the leadup to the UK summit this week - an official on a public briefing just now called this "the most significant action any government has ever taken on AI," and the factsheet says ⬇️

Tweet media one

2

6

50

@hlntnr

Helen Toner

11 months

The caveat: In the piece we breeze past the question of whether regulating AI would actually slow down US innovation. "We can't regulate because China" assumes that regs = slowdown, but this is far from certain! Smart regulation can be neutral or even positive for innovation.

4

4

50

@hlntnr

Helen Toner

4 years

Dispatches from life in the natsec & tech space: was in a workshop today with more men named Ryan than women. (there were FOUR RYANS, how is that even possible)

4

1

51