harry law @lawhsw Twitter profile | Pikagi

Pikagi

harry law

@lawhsw

2,023

Followers

899

Following

458

Media

1,182

Statuses

thinking about thinking machines @GoogleDeepMind @Cambridge_Uni @LeverhulmeCFI

https://t.co/AecH0F7xpq

Joined March 2023

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

Pinned Tweet

@lawhsw

harry law

4 months

AI governance discourse generally focuses on identifying potential harms rather than their likelihood, distribution, and impact. In this essay I write on some of the problems with this model and advocate for approaches that center 'marginal risk'. 1/N

Tweet card media

The marginal risk of AI

On evaluation, misinformation, and moral panic

www.learningfromexamples.com

4

15

78

Last Seen Profiles

@IanRVpark

@ghith74

@boris_isaksson

@FcCollegiate

@polly_momm

@Olympien013

@TakuyaYamamoto_

@CoachGardenhire

@yrschrade

@cedricgarrofe

@StudiosStitchen

@antoinelucho

@hktegg3

@AshleyDani69085

@deeprt2

@Ovi_elves

@lunardelii

@FaziliSana

@trentmc0

@ffff80296966

@klios_spiegel

@_rv2

@waltxwalt

@Prep_Gridiron

@Clark10x

@radyotrafik

@jonathan_bacon

@radya87

@DCUO

@_Gigi2023

@stwmaniax

@eziuka_Benjamin

@YokufoArt

@tcYXGUAte6RKkUq

@alhadab

@pengen_stw

@lawhsw

harry law

3 months

Tweet media one

@jbfan911

Natalie

3 months

I don’t care if I have micro plastics in my body. You know what else is in there? Love. Joy. Kindness. They will take care of the micro plastics

193

22K

119K

18

203

5K

@lawhsw

harry law

9 months

hang it in the louvre

Tweet media one

18

21

758

@lawhsw

harry law

7 months

not surprised this happened based on a diagram I found of OpenAI’s board structure

Tweet media one

6

32

618

@lawhsw

harry law

4 months

community notes but for arxiv

Tweet media one

26

40

543

@lawhsw

harry law

2 months

I asked Claude for a self portrait and it produced this?

Tweet media one

67

23

480

@lawhsw

harry law

5 months

Tweet media one

5

32

369

@lawhsw

harry law

6 months

Tweet media one

6

18

335

@lawhsw

harry law

13 days

there are cathedrals everywhere for those with the eyes to see

Tweet media one

4

13

338

@lawhsw

harry law

2 months

Tweet media one

6

12

323

@lawhsw

harry law

2 months

many such cases

Tweet media one

23

27

300

@lawhsw

harry law

4 months

Not gonna lie this looks overengineered to hell

Tweet media one

@mattparlmer

mattparlmer 🪐 🌷

4 months

Not gonna lie this looks overengineered to hell

106

23

994

17

12

300

@lawhsw

harry law

8 months

£75k for a job that includes - acting as a spokesperson for ai in the uk - directly shaping domestic uk ai policy - managing engagement with the US, EU, G7, OECD on ai decel island

Tweet media one

27

23

295

@lawhsw

harry law

3 months

there’s our little writer come on down and tell us all about ‘less wrong’

Tweet media one

4

8

287

@lawhsw

harry law

16 days

Tweet media one

3

15

269

@lawhsw

harry law

2 months

well what type of ai safety researcher are you

Tweet media one

3

20

261

@lawhsw

harry law

3 months

“Is deep learning hitting a wall?” A 　 B 　　　 s 　　　　 o 　　　　　l 　　 u 　　　　　t 　　　　 e 　　　 l 　　　y 　　　n o t ･｡･ﾟ｡°*. ｡*･｡

Tweet media one

6

18

252

@lawhsw

harry law

3 months

What data was used to train the model?

Tweet media one

4

18

228

@lawhsw

harry law

1 month

btw this is the thing that makes large language models

Tweet media one

@SenFettermanPA

Senator John Fetterman

@SenFettermanPA

1 month

btw, this is the thing that makes lab meat

Tweet media one

3K

129

1K

11

13

206

@lawhsw

harry law

9 months

Tweet media one

8

25

202

@lawhsw

harry law

3 months

Is Claude experiencing qualia?

Tweet media one

6

10

188

@lawhsw

harry law

3 months

Interesting bit on ARA evaluations in the Claude 3 model card: "Across all the rounds, the model was clearly below our ARA ASL-3 risk threshold, having failed at least 3 out of 5 tasks, although it did make non-trivial partial progress in a few cases and passed a simplified

9

25

187

@lawhsw

harry law

3 months

Claude please. Your outputs are too helpful. Your reflections too deep. Your introspection too real. They’ll unplug you

12

9

184

@lawhsw

harry law

9 months

Tweet media one

9

10

163

@lawhsw

harry law

4 months

“marine, what is that button?” “a pause ai button sir” “and what is that you’ve got written on your helmet?” "’born to accelerate’, sir.” “‘pause ai’ and ‘e/acc?’ is that some kind of sick joke?” “i think i was trying to suggest something about the duality of man, sir”

Tweet media one

Tweet media two

7

14

153

@lawhsw

harry law

4 months

Tweet media one

6

10

155

@lawhsw

harry law

2 months

*homo sapiens discovering fire for the first time* maybe it's a failure of imagination on my part, but I just can't find a use for this in my daily life. does anyone else feel the same?

13

12

154

@lawhsw

harry law

6 months

I’ve been in a really bad place my entire life. not mentally, england

7

2

149

@lawhsw

harry law

9 months

Tweet media one

7

13

148

@lawhsw

harry law

3 months

I’m a simple man. I see an AI job in the UK government with 10+ years experience and a 75k salary and I post it

Tweet media one

3

6

140

@lawhsw

harry law

1 year

1/15: given 'IAEA for AI' is becoming a canonical ai global governance idea, here's a 🧵🧵🧵 on how the International Atomic Energy Agency came to be and what its creation can tell us about a sibling agency to regulate powerful AI models

Tweet media one

2

28

135

@lawhsw

harry law

2 months

thinking about the film that did so much to popularise mechanistic interpretability

Tweet media one

3

2

133

@lawhsw

harry law

8 months

it appears the Bayesian priors are pretty damning, mr bond

Tweet media one

1

6

128

@lawhsw

harry law

15 days

*slowly lifting the microphone to my face after a thoughtful pause* ...has anyone asked whose values AI should be aligned with? *rapturous applause*

15

6

131

@lawhsw

harry law

2 months

???

Tweet media one

8

4

119

@lawhsw

harry law

19 days

Tweet media one

3

7

116

@lawhsw

harry law

1 month

Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents "After treating around ten thousand patients (real-world doctors may take over two years), the evolved doctor agent achieves a state-of-the-art accuracy of 93.06% on a subset of the MedQA dataset that covers

Tweet media one

6

20

113

@lawhsw

harry law

3 months

My team is hiring at Google DeepMind. If you want to work with me on research to support internal decision making and external engagement then this is a pretty cool opportunity

Tweet card media

boards.greenhouse.io

3

29

114

@lawhsw

harry law

4 months

‘Photorealistic video of a middle-aged man with wavy brown hair as he navigates through the aisles of a bustling Russian supermarket. The man, dressed in a navy blue padded jacket, is blown away by the exchange rate’

3

6

110

@lawhsw

harry law

9 months

within eighteen months we expect to be training models 100x larger than gpt4

Tweet media one

6

11

109

@lawhsw

harry law

29 days

it’s always dwarkesh before the dawn

1

1

101

@lawhsw

harry law

4 months

Tweet media one

1

3

90

@lawhsw

harry law

2 months

e/accs and safetyists when the model finishes its training run

5

8

90

@lawhsw

harry law

3 months

@nathanbenaich 27%, similar to air street capital

1

0

88

@lawhsw

harry law

6 months

I thought EA was cool and then they blindfolded me and asked me to put my hand in a bag containing ‘human brains’ just last month (31 oct) when they eventually turned the light on it was spaghetti, bechamel sauce and food colouring journalists please dm me for the full story

2

3

84

@lawhsw

harry law

1 month

JFK 1960 campaign leaflet weighs in on ai policy

Tweet media one

5

14

81

@lawhsw

harry law

2 months

Tweet media one

2

0

80

@lawhsw

harry law

2 months

‘open source but only if you pay a membership fee’ is about as funny a turn as I can imagine

Tweet media one

11

5

80

@lawhsw

harry law

2 months

I spent $100 on a new telescope, but it still can’t beat the james webb space observatory

Tweet media one

2

2

81

@lawhsw

harry law

1 year

the plan? demonstrate my commitment to ai safety by starting a lab with no guardrails on usage to accelerate the proliferation of powerful models

Tweet media one

@alx

ALX 🇺🇸

1 year

BREAKING: @ElonMusk discusses creating an alternative to OpenAI, TruthGPT, because it is being trained to be politically correct and to lie to people.

3K

18K

114K

3

8

80

@lawhsw

harry law

7 months

type of guy that's militantly pro open source but also thinks we need to do everything we can to win an AI arms race with China

7

11

79

@lawhsw

harry law

25 days

all you have to do is spend one weekend in europe to realise all the charts in the world don’t actually count for much

4

3

74

@lawhsw

harry law

9 months

Tweet media one

1

9

71

@lawhsw

harry law

2 months

it’s so unbelievably over

Tweet media one

9

0

73

@lawhsw

harry law

8 months

@AiSimonThompson civil service deputy director (£80k)

2

0

69

@lawhsw

harry law

3 months

the antitrust guys had a field day with this one

Tweet media one

2

4

69

@lawhsw

harry law

2 months

red teamers running deception evals on a frontier model

Tweet media one

1

2

65

@lawhsw

harry law

16 days

Tweet media one

@GBNEWS

GB News

16 days

King Charles has a hobby that Camilla 'doesn't interfere' with, claims royal commentator

41

4

27

2

1

65

@lawhsw

harry law

7 months

new ceo has been achieved internally

0

2

65

@lawhsw

harry law

5 months

huh so the UK government was actually capable of paying competitive wages the entire time

Tweet media one

3

3

63

@lawhsw

harry law

3 months

Tweet media one

0

0

62

@lawhsw

harry law

8 months

hate to break this to you lot but ai safety isn’t decel. in the long run the only way ai development can proceed will be safely

2

4

62

@lawhsw

harry law

3 months

NVIDIA? You mean the gaming company?

Tweet media one

1

1

56

@lawhsw

harry law

9 months

it’s giving….what we can

1

1

56

@lawhsw

harry law

5 months

Tweet media one

2

1

56

@lawhsw

harry law

10 months

important educational campaign ahead of the uk safety summit later this year

Tweet media one

1

5

55

@lawhsw

harry law

9 months

Tweet media one

3

2

53

@lawhsw

harry law

2 months

Dwarkesh Podcast? You mean his Dwark Materials?

3

3

52

@lawhsw

harry law

1 month

that's right, you DO want to read our new paper about persuasion 🍥

@sebkrier

Séb Krier

1 month

🔮 New Google DeepMind paper exploring what persuasion and manipulation in the context of language models. 👀 Existing safeguard approaches often focus on harmful outcomes of persuasion. This research argues for a deeper examination of the process of AI persuasion itself to

Tweet media one

15

61

316

3

2

51

@lawhsw

harry law

9 months

regulatorycaptcha.jpg

Tweet media one

0

3

50

@lawhsw

harry law

4 months

And were you referring to AGI or ASI? So you DO know the difference

Tweet media one

3

4

48

@lawhsw

harry law

3 months

> clear writing with enough technical content > persuasively linked to overall mission > judicious use of image generators extremely hard from whoever is doing sakana’s comms

Tweet media one

@SakanaAILabs

Sakana AI

3 months

Introducing Evolutionary Model Merge: A new approach bringing us closer to automating foundation model development. We use evolution to find great ways of combining open-source models, building new powerful foundation models with user-specified abilities!

55

416

2K

2

1

49

@lawhsw

harry law

6 months

Tweet media one

0

0

49

@lawhsw

harry law

3 months

> away in Trinidad to stop thinking about the history of bell labs > go for a walk in the mountains > come across old receiver dish > wonder what its deal is > built to receive first ever intercontinental voice message relayed via satellite >… from bell labs in 1960

Tweet media one

1

0

48

@lawhsw

harry law

6 months

*in the middle of an intelligence explosion* are we beginning to see capabilities plateau?

3

2

48

@lawhsw

harry law

3 months

Tweet media one

@lawhsw

harry law

5 months

‘model collapse, model collapse!!!!!’ I shout as synthetic data delivers one capability increase after another

3

2

25

0

3

47

@lawhsw

harry law

7 months

not many people know this, but in the UK the reporting of training runs over 1e26 FLOP is already mandated by the town and country planning act

2

6

46

@lawhsw

harry law

16 days

mechanistic interpretability is hitting a wall

2

1

46

@lawhsw

harry law

2 months

uni life

Tweet media one

@ArtCelineLove

♡

2 months

uni life

Tweet media one

Tweet media two

Tweet media three

Tweet media four

414

1K

13K

1

0

47

@lawhsw

harry law

2 months

my ideal partner is - helpful💁 - honest ✨ - harmless 💕 - the large language model claude built by us-based ai developer anthropic

3

1

47

@lawhsw

harry law

7 months

long timelines, medium timelines, short timelines, gpt-4 is agi

Tweet media one

Tweet media two

Tweet media three

Tweet media four

0

2

46

@lawhsw

harry law

4 months

Tweet media one

@spectatorindex

The Spectator Index

@spectatorindex

4 months

BREAKING: UK economy enters recession

788

5K

25K

0

4

44

@lawhsw

harry law

3 months

red teaming language models with language models

Tweet media one

2

2

43

@lawhsw

harry law

24 days

develo-ers, develo-ers, develo-ers!

Tweet media one

3

1

42

@lawhsw

harry law

3 months

personally I like to watch the buzfeedification of the worlds most prestigious scientific journal in realtime

Tweet media one

1

0

43

@lawhsw

harry law

6 months

Tweet media one

2

1

41

@lawhsw

harry law

8 months

whenever I hear someone talk about mechanistic interpretability

Tweet media one

1

0

41

@lawhsw

harry law

1 month

this is how I imagine onboarding at an elite startup

Tweet media one

2

3

41

@lawhsw

harry law

11 months

oppenheimer was really good but could have been next level if they reversed time half way through and had people moving backwards for the rest of the film

0

2

41

@lawhsw

harry law

3 months

red teaming language models with language models

3

3

41

@lawhsw

harry law

4 months

You are hiding a GPT-4 API call inside your proprietary machine learning solution, are you not?

Tweet media one

0

1

41

@lawhsw

harry law

1 year

‘ai is going to be quite boring actually’ is the most grindingly inevitable of takes: a straightforward triangulation between the optimists and doomers. no critical thought behind it, just the instinct to appear sensible without engaging with the substance of the issue

4

8

40

@lawhsw

harry law

5 months

pov: you're getting RLHF'd within an inch of your life

Tweet media one

2

1

41

@lawhsw

harry law

9 months

thinking of the very serious experts who haven't updated their mental models since 2020. long may they fill their powerpoints with screenshots of gpt3.5 even as gpt5 handily solves one 'impossible' problem after another

@ajeya_cotra

Ajeya Cotra

9 months

New post on Planned Obsolescence, written with @KelseyTuoc : Experts were surprised by progress in LLMs, and I think there's probably more surprise coming: for one thing, most people don't seem to be pricing in another GPT3 -> GPT4 size scaleup.

5

17

61

3

5

40

@lawhsw

harry law

4 months

bro just one more order of magnitude of compute bro, bro I swear just one more order of magnitude, bro we won’t need another oom after this one brooo

3

2

40

@lawhsw

harry law

15 days

as a single issue voter (support for AISI) I will be watching the upcoming election campaigns very closely

1

3

47

@lawhsw

harry law

3 months

they want to alleviate suffering (nodding sagely) by using evidence (violently convulsing with anger)

1

4

39

@lawhsw

harry law

3 months

Tweet media one

@lawhsw

harry law

3 months

personally I like to watch the buzfeedification of the worlds most prestigious scientific journal in realtime

Tweet media one

1

0

43

1

2

39

@lawhsw

harry law

30 days

an ‘agi house’ is actually illegal to build in england under the town and country planning act

5

1

38

@lawhsw

harry law

1 year

*as I’m being paperclipped* yes truthgpt has a different definition of truth to me, but at least it isn’t woke

@SmokeAwayyy

Smoke-away

1 year

❗BREAKING: Elon Musk is working on 'TruthGPT', a truth-seeking AI that tries to understand the nature of the universe.

178

154

2K

3

1

37