RSS

AI Capabilities

TagLast edit: 29 Aug 2021 12:57 UTC by plex

AI Capabilities are the growing abilities of AIs to act effectively in increasingly complex environments. It is often compared to to AI Alignment, which refers to efforts to ensure that these effective actions taken by AIs are also intended by the creators and beneficial to humanity.

Effi­cien­tZero: hu­man ALE sam­ple-effi­ciency w/​MuZero+self-supervised

gwern2 Nov 2021 2:32 UTC
137 points
52 comments1 min readLW link
(arxiv.org)

A small up­date to the Sparse Cod­ing in­terim re­search report

30 Apr 2023 19:54 UTC
61 points
5 comments1 min readLW link

[Paper] Stress-test­ing ca­pa­bil­ity elic­i­ta­tion with pass­word-locked models

4 Jun 2024 14:52 UTC
84 points
10 comments12 min readLW link
(arxiv.org)

Me­moriz­ing weak ex­am­ples can elicit strong be­hav­ior out of pass­word-locked models

6 Jun 2024 23:54 UTC
58 points
5 comments7 min readLW link

Get­ting 50% (SoTA) on ARC-AGI with GPT-4o

ryan_greenblatt17 Jun 2024 18:44 UTC
262 points
49 comments13 min readLW link

Com­pet­i­tive pro­gram­ming with AlphaCode

Algon2 Feb 2022 16:49 UTC
58 points
36 comments15 min readLW link
(deepmind.com)

Effi­cien­tZero: How It Works

1a3orn26 Nov 2021 15:17 UTC
297 points
50 comments29 min readLW link1 review

Meta AI an­nounces Cicero: Hu­man-Level Di­plo­macy play (with di­alogue)

Jacy Reese Anthis22 Nov 2022 16:50 UTC
93 points
64 comments1 min readLW link
(www.science.org)

Deep­Mind on Strat­ego, an im­perfect in­for­ma­tion game

sanxiyn24 Oct 2022 5:57 UTC
15 points
9 comments1 min readLW link
(arxiv.org)

What will the scaled up GATO look like? (Up­dated with ques­tions)

Amal 25 Oct 2022 12:44 UTC
34 points
22 comments1 min readLW link

[Question] The thing I don’t un­der­stand about AGI

Jeremy Kalfus18 Jun 2024 4:25 UTC
7 points
12 comments1 min readLW link

Devil’s Ad­vo­cate: Ad­verse Selec­tion Against Con­scien­tious­ness

lionhearted (Sebastian Marshall)28 May 2023 17:53 UTC
10 points
2 comments1 min readLW link

Is AI Progress Im­pos­si­ble To Pre­dict?

alyssavance15 May 2022 18:30 UTC
277 points
39 comments2 min readLW link

[Cross­post] AlphaTen­sor, Taste, and the Scal­a­bil­ity of AI

jamierumbelow9 Oct 2022 19:42 UTC
16 points
4 comments1 min readLW link
(jamieonsoftware.com)

What DALL-E 2 can and can­not do

Swimmer963 (Miranda Dixon-Luinenburg) 1 May 2022 23:51 UTC
353 points
303 comments9 min readLW link

The case for a nega­tive al­ign­ment tax

18 Sep 2024 18:33 UTC
79 points
20 comments7 min readLW link

[linkpost] The fi­nal AI bench­mark: BIG-bench

RomanS10 Jun 2022 8:53 UTC
25 points
21 comments1 min readLW link

Timelines to Trans­for­ma­tive AI: an investigation

Zershaaneh Qureshi26 Mar 2024 18:28 UTC
20 points
2 comments50 min readLW link

AlphaGeom­e­try: An Olympiad-level AI sys­tem for geometry

alyssavance17 Jan 2024 17:17 UTC
45 points
9 comments1 min readLW link
(deepmind.google)

Ca­pa­bil­ities and al­ign­ment of LLM cog­ni­tive architectures

Seth Herd18 Apr 2023 16:29 UTC
86 points
18 comments20 min readLW link

Prin­ci­ples of Pri­vacy for Align­ment Research

johnswentworth27 Jul 2022 19:53 UTC
72 points
31 comments7 min readLW link

The longest train­ing run

17 Aug 2022 17:18 UTC
71 points
12 comments9 min readLW link
(epochai.org)

[Question] Are lan­guage mod­els close to the su­per­hu­man level in philos­o­phy?

Roman Leventov19 Aug 2022 4:43 UTC
6 points
2 comments2 min readLW link

What’s the Most Im­pres­sive Thing That GPT-4 Could Plau­si­bly Do?

bayesed26 Aug 2022 15:34 UTC
24 points
22 comments1 min readLW link

[Question] What would you ex­pect a mas­sive mul­ti­modal on­line fed­er­ated learner to be ca­pa­ble of?

Aryeh Englander27 Aug 2022 17:31 UTC
13 points
4 comments1 min readLW link

Read­abil­ity is mostly a waste of characters

vlad.proex21 Apr 2023 22:05 UTC
21 points
7 comments3 min readLW link

No, hu­man brains are not (much) more effi­cient than computers

Jesse Hoogland6 Sep 2022 13:53 UTC
22 points
21 comments3 min readLW link
(www.jessehoogland.com)

Alex­aTM − 20 Billion Pa­ram­e­ter Model With Im­pres­sive Performance

MrThink9 Sep 2022 21:46 UTC
5 points
0 comments1 min readLW link

Eval­u­a­tions pro­ject @ ARC is hiring a re­searcher and a web­dev/​engineer

Beth Barnes9 Sep 2022 22:46 UTC
99 points
7 comments10 min readLW link

[Question] Are Speed Su­per­in­tel­li­gences Fea­si­ble for Modern ML Tech­niques?

DragonGod14 Sep 2022 12:59 UTC
9 points
7 comments1 min readLW link

Steer­ing sub­sys­tems: ca­pa­bil­ities, agency, and alignment

Seth Herd29 Sep 2023 13:45 UTC
26 points
0 comments8 min readLW link

ACT-1: Trans­former for Actions

Daniel Kokotajlo14 Sep 2022 19:09 UTC
52 points
4 comments1 min readLW link
(www.adept.ai)

[Question] Could trans­former net­work mod­els learn mo­tor plan­ning like they can learn lan­guage and image gen­er­a­tion?

mu_(negative)23 Apr 2023 17:24 UTC
2 points
4 comments1 min readLW link

Molec­u­lar dy­nam­ics data will be es­sen­tial for the next gen­er­a­tion of ML pro­tein models

Abhishaike Mahajan26 Aug 2024 14:50 UTC
9 points
0 comments11 min readLW link
(www.owlposting.com)

The Stochas­tic Par­rot Hy­poth­e­sis is de­bat­able for the last gen­er­a­tion of LLMs

7 Nov 2023 16:12 UTC
52 points
20 comments6 min readLW link

Will we run out of ML data? Ev­i­dence from pro­ject­ing dataset size trends

Pablo Villalobos14 Nov 2022 16:42 UTC
75 points
12 comments2 min readLW link
(epochai.org)

Mas­ter­ing Strat­ego (Deep­mind)

svemirski2 Dec 2022 2:21 UTC
6 points
0 comments1 min readLW link
(www.deepmind.com)

Can GPT-3 Write Con­tra Dances?

jefftk4 Dec 2022 3:00 UTC
6 points
4 comments10 min readLW link
(www.jefftk.com)

A Year of AI In­creas­ing AI Progress

TW12330 Dec 2022 2:09 UTC
148 points
3 comments2 min readLW link

[Question] Is “Re­cur­sive Self-Im­prove­ment” Rele­vant in the Deep Learn­ing Paradigm?

DragonGod6 Apr 2023 7:13 UTC
32 points
36 comments7 min readLW link

Lan­guage mod­els can gen­er­ate su­pe­rior text com­pared to their input

ChristianKl17 Jan 2023 10:57 UTC
48 points
28 comments1 min readLW link

Google an­nounces ‘Bard’ pow­ered by LaMDA

M. Y. Zuo6 Feb 2023 19:40 UTC
31 points
3 comments2 min readLW link

Syd­ney can play chess and kind of keep track of the board state

Erik Jenner3 Mar 2023 9:39 UTC
64 points
19 comments6 min readLW link

Google’s PaLM-E: An Em­bod­ied Mul­ti­modal Lan­guage Model

SandXbox7 Mar 2023 4:11 UTC
87 points
7 comments1 min readLW link
(palm-e.github.io)

Squeez­ing foun­da­tions re­search as­sis­tance out of for­mal logic nar­row AI.

Donald Hobson8 Mar 2023 9:38 UTC
16 points
1 comment2 min readLW link

A chess game against GPT-4

Rafael Harth16 Mar 2023 14:05 UTC
24 points
23 comments1 min readLW link

Why the tech­nolog­i­cal sin­gu­lar­ity by AGI may never happen

hippke3 Sep 2021 14:19 UTC
5 points
14 comments1 min readLW link

Epistemic Strate­gies of Safety-Ca­pa­bil­ities Tradeoffs

adamShimi22 Oct 2021 8:22 UTC
5 points
0 comments6 min readLW link

The al­ign­ment prob­lem in differ­ent ca­pa­bil­ity regimes

Buck9 Sep 2021 19:46 UTC
88 points
12 comments5 min readLW link

Google an­nounces Path­ways: new gen­er­a­tion mul­ti­task AI Architecture

Ozyrus29 Oct 2021 11:55 UTC
6 points
1 comment1 min readLW link
(blog.google)

Bench­mark­ing LLM Agents on Kag­gle Competitions

aogara22 Mar 2024 13:09 UTC
15 points
4 comments5 min readLW link

“AI achieves silver-medal stan­dard solv­ing In­ter­na­tional Math­e­mat­i­cal Olympiad prob­lems”

gjm25 Jul 2024 15:58 UTC
133 points
38 comments2 min readLW link
(deepmind.google)

Diffu­sion Guided NLP: bet­ter steer­ing, mostly a good thing

Nathan Helm-Burger10 Aug 2024 19:49 UTC
13 points
0 comments1 min readLW link
(arxiv.org)

In­ter­pret­ing Yud­kowsky on Deep vs Shal­low Knowledge

adamShimi5 Dec 2021 17:32 UTC
100 points
32 comments24 min readLW link

Re­quest: stop ad­vanc­ing AI capabilities

So8res26 May 2023 17:42 UTC
153 points
24 comments1 min readLW link

AI do­ing philos­o­phy = AI gen­er­at­ing hands?

Wei Dai15 Jan 2024 9:04 UTC
46 points
22 comments1 min readLW link

OpenAI Solves (Some) For­mal Math Olympiad Problems

Michaël Trazzi2 Feb 2022 21:49 UTC
78 points
27 comments2 min readLW link

[Question] Killing Re­cur­rent Me­mory Over Self At­ten­tion?

Del Nobolo6 Jun 2023 23:02 UTC
3 points
0 comments1 min readLW link

Per­sonal imi­ta­tion software

Flaglandbase7 Mar 2022 7:55 UTC
6 points
6 comments1 min readLW link

Elon Musk an­nounces xAI

Jan_Kulveit13 Jul 2023 9:01 UTC
75 points
35 comments1 min readLW link
(www.ft.com)

ChatGPT and Bing Chat can’t play Botticelli

Asha Saavoss29 Mar 2023 17:39 UTC
11 points
0 comments6 min readLW link

PaLM in “Ex­trap­o­lat­ing GPT-N perfor­mance”

Lukas Finnveden6 Apr 2022 13:05 UTC
83 points
19 comments2 min readLW link

Dual-Use­ness is a Ratio

jimrandomh6 Apr 2023 5:46 UTC
35 points
2 comments1 min readLW link

We have achieved Noob Gains in AI

phdead18 May 2022 20:56 UTC
117 points
20 comments7 min readLW link

Un­com­pet­i­tive pro­gram­ming with GPT-3

Bezzi6 Feb 2022 10:19 UTC
7 points
8 comments3 min readLW link

$300 for the best sci-fi prompt: the results

RomanS3 Jan 2024 19:10 UTC
16 points
19 comments7 min readLW link

Ques­tions I’d Want to Ask an AGI+ to Test Its Un­der­stand­ing of Ethics

sweenesm26 Jan 2024 23:40 UTC
14 points
6 comments4 min readLW link

On agen­tic gen­er­al­ist mod­els: we’re es­sen­tially us­ing ex­ist­ing tech­nol­ogy the weak­est and worst way you can use it

Yuli_Ban28 Aug 2024 1:57 UTC
10 points
2 comments9 min readLW link

An In­tro­duc­tion to AI Sandbagging

26 Apr 2024 13:40 UTC
44 points
10 comments8 min readLW link

[Paper] AI Sand­bag­ging: Lan­guage Models can Strate­gi­cally Un­der­perform on Evaluations

13 Jun 2024 10:04 UTC
84 points
10 comments2 min readLW link
(arxiv.org)

What’s the fu­ture of AI hard­ware?

Itay Dreyfus17 Jun 2024 13:05 UTC
2 points
0 comments8 min readLW link
(productidentity.co)

A short pro­ject on Mamba: grokking & interpretability

Alejandro Tlaie18 Oct 2024 16:59 UTC
21 points
0 comments6 min readLW link

Agen­tized LLMs will change the al­ign­ment landscape

Seth Herd9 Apr 2023 2:29 UTC
157 points
97 comments3 min readLW link

Sta­bil­ity AI re­leases StableLM, an open-source ChatGPT counterpart

Ozyrus20 Apr 2023 6:04 UTC
11 points
3 comments1 min readLW link
(github.com)

[Thought Ex­per­i­ment] To­mor­row’s Echo—The fu­ture of syn­thetic com­pan­ion­ship.

Vimal Naran26 Oct 2023 17:54 UTC
−7 points
2 comments2 min readLW link

AI as Su­per-Demagogue

RationalDino5 Nov 2023 21:21 UTC
0 points
11 comments9 min readLW link

A call for a quan­ti­ta­tive re­port card for AI bioter­ror­ism threat models

Juno4 Dec 2023 6:35 UTC
12 points
0 comments10 min readLW link

GPT4 is ca­pa­ble of writ­ing de­cent long-form sci­ence fic­tion (with the right prompts)

RomanS23 May 2023 13:41 UTC
22 points
28 comments65 min readLW link

AGI-Au­to­mated In­ter­pretabil­ity is Suicide

__RicG__10 May 2023 14:20 UTC
23 points
33 comments7 min readLW link

GPT-4 im­plic­itly val­ues iden­tity preser­va­tion: a study of LMCA iden­tity management

Ozyrus17 May 2023 14:13 UTC
21 points
4 comments13 min readLW link

TinyS­to­ries: Small Lan­guage Models That Still Speak Co­her­ent English

Ulisse Mini28 May 2023 22:23 UTC
66 points
8 comments2 min readLW link
(arxiv.org)

[Question] Hy­po­thet­i­cal: what would you do?

JNS3 Aug 2023 22:39 UTC
4 points
2 comments1 min readLW link

LLMs are (mostly) not helped by filler tokens

Kshitij Sachan10 Aug 2023 0:48 UTC
66 points
35 comments6 min readLW link

In­flec­tion.ai is a ma­jor AGI lab

nikola9 Aug 2023 1:05 UTC
137 points
13 comments2 min readLW link

Google Deep­Mind’s RT-2

SandXbox11 Aug 2023 11:26 UTC
9 points
1 comment1 min readLW link
(robotics-transformer2.github.io)

Stu­pidity is also hard

walkthroughwalls12 Sep 2023 2:45 UTC
−8 points
4 comments2 min readLW link

Ba­sic Math­e­mat­ics of Pre­dic­tive Coding

Adam Shai29 Sep 2023 14:38 UTC
49 points
6 comments9 min readLW link

Towards Bet­ter Mile­stones for Mon­i­tor­ing AI Capabilities

snewman27 Sep 2023 21:18 UTC
11 points
0 comments14 min readLW link

[Question] Is there a pub­li­cly available list of ex­am­ples of fron­tier model ca­pa­bil­ities?

Max Kearney19 Sep 2023 17:45 UTC
1 point
0 comments1 min readLW link

In­ter­pretabil­ity Ex­ter­nal­ities Case Study—Hun­gry Hun­gry Hippos

Magdalena Wache20 Sep 2023 14:42 UTC
64 points
22 comments2 min readLW link

This anime sto­ry­board doesn’t ex­ist: a graphic novel writ­ten and illus­trated by GPT4

RomanS5 Oct 2023 14:01 UTC
12 points
7 comments55 min readLW link

I Would Have Solved Align­ment, But I Was Wor­ried That Would Ad­vance Timelines

307th20 Oct 2023 16:37 UTC
118 points
33 comments9 min readLW link

Eleuther re­leases Llemma: An Open Lan­guage Model For Mathematics

mako yass17 Oct 2023 20:03 UTC
22 points
0 comments1 min readLW link
(blog.eleuther.ai)

[Question] What are the rel­a­tive speeds of AI ca­pa­bil­ities and AI safety?

NunoSempere24 Apr 2020 18:21 UTC
8 points
2 comments1 min readLW link

Deep­Mind: Gen­er­ally ca­pa­ble agents emerge from open-ended play

Daniel Kokotajlo27 Jul 2021 14:19 UTC
247 points
53 comments2 min readLW link
(deepmind.com)

OpenAI Codex: First Impressions

specbug13 Aug 2021 16:52 UTC
49 points
8 comments4 min readLW link
(sixeleven.in)

To con­tribute to AI safety, con­sider do­ing AI research

Vika16 Jan 2016 20:42 UTC
39 points
39 comments2 min readLW link

[Question] What’s the differ­ence be­tween newer Atari-play­ing AI and the older Deep­mind one (from 2014)?

Raemon2 Nov 2021 23:36 UTC
27 points
8 comments1 min readLW link

AI Tracker: mon­i­tor­ing cur­rent and near-fu­ture risks from su­per­scale models

23 Nov 2021 19:16 UTC
67 points
13 comments3 min readLW link
(aitracker.org)

HIRING: In­form and shape a new pro­ject on AI safety at Part­ner­ship on AI

Madhulika Srikumar24 Nov 2021 8:27 UTC
6 points
0 comments1 min readLW link

How to mea­sure FLOP/​s for Neu­ral Net­works em­piri­cally?

Marius Hobbhahn29 Nov 2021 15:18 UTC
16 points
5 comments7 min readLW link

What’s the back­ward-for­ward FLOP ra­tio for Neu­ral Net­works?

13 Dec 2021 8:54 UTC
20 points
12 comments10 min readLW link

How I’m think­ing about GPT-N

delton13717 Jan 2022 17:11 UTC
54 points
21 comments18 min readLW link

Es­ti­mat­ing train­ing com­pute of Deep Learn­ing models

20 Jan 2022 16:12 UTC
37 points
4 comments1 min readLW link

Lifel­og­ging for Align­ment & Immortality

Dev.Errata17 Aug 2024 23:42 UTC
13 points
3 comments7 min readLW link

Test­ing PaLM prompts on GPT3

Yitz6 Apr 2022 5:21 UTC
103 points
14 comments8 min readLW link

Gato’s Gen­er­al­i­sa­tion: Pre­dic­tions and Ex­per­i­ments I’d Like to See

Oliver Sourbut18 May 2022 7:15 UTC
43 points
3 comments10 min readLW link

[Question] What is the most prob­a­ble AI?

Zeruel01720 Jun 2022 23:26 UTC
−2 points
0 comments3 min readLW link

AI Fore­cast­ing: One Year In

jsteinhardt4 Jul 2022 5:10 UTC
132 points
12 comments6 min readLW link
(bounded-regret.ghost.io)

A Cri­tique of AI Align­ment Pessimism

ExCeph19 Jul 2022 2:28 UTC
9 points
1 comment9 min readLW link

Align­ment be­ing im­pos­si­ble might be bet­ter than it be­ing re­ally difficult

Martín Soto25 Jul 2022 23:57 UTC
13 points
2 comments2 min readLW link

[Question] How might we make bet­ter use of AI ca­pa­bil­ities re­search for al­ign­ment pur­poses?

Jemal Young31 Aug 2022 4:19 UTC
11 points
4 comments1 min readLW link

How should Deep­Mind’s Chin­chilla re­vise our AI fore­casts?

Cleo Nardo15 Sep 2022 17:54 UTC
35 points
12 comments13 min readLW link

It mat­ters when the first sharp left turn happens

Adam Jermyn29 Sep 2022 20:12 UTC
44 points
9 comments4 min readLW link

Anony­mous ad­vice: If you want to re­duce AI risk, should you take roles that ad­vance AI ca­pa­bil­ities?

Benjamin Hilton11 Oct 2022 14:16 UTC
54 points
9 comments1 min readLW link

Is GPT-N bounded by hu­man ca­pa­bil­ities? No.

Cleo Nardo17 Oct 2022 23:26 UTC
48 points
8 comments2 min readLW link

They gave LLMs ac­cess to physics simulators

ryan_b17 Oct 2022 21:21 UTC
50 points
18 comments1 min readLW link
(arxiv.org)

Ar­ti­cle Re­view: Google’s AlphaTensor

Robert_AIZI12 Oct 2022 18:04 UTC
8 points
4 comments10 min readLW link

Paper: Dis­cov­er­ing novel al­gorithms with AlphaTen­sor [Deep­mind]

LawrenceC5 Oct 2022 16:20 UTC
82 points
18 comments1 min readLW link
(www.deepmind.com)

[Question] Is the speed of train­ing large mod­els go­ing to in­crease sig­nifi­cantly in the near fu­ture due to Cere­bras An­dromeda?

Amal 15 Nov 2022 22:50 UTC
13 points
11 comments1 min readLW link

When AI solves a game, fo­cus on the game’s me­chan­ics, not its theme.

Cleo Nardo23 Nov 2022 19:16 UTC
88 points
7 comments2 min readLW link

Notes on Meta’s Di­plo­macy-Play­ing AI

Erich_Grunewald22 Dec 2022 11:34 UTC
14 points
2 comments14 min readLW link
(www.erichgrunewald.com)

A case for ca­pa­bil­ities work on AI as net pos­i­tive

Noosphere8927 Feb 2023 21:12 UTC
10 points
37 comments1 min readLW link