GaryMarcus Profile Banner
Gary Marcus Profile
Gary Marcus

@GaryMarcus

Followers
182K
Following
78K
Media
3K
Statuses
51K

Built two AI companies, wrote six books, tried to warn you about a lot of things.

Joined December 2010
Don't wanna be here? Send us removal request.
@GaryMarcus
Gary Marcus
7 days
Nine June 2029 AI predictions — from someone who was basically 7/7 on 2024:. 1. The LLM race will be basically a tie between China and U.S. 2. Pure LLMs will still hallucinate (alternative architectures may do better). 3. Pure LLMs will still make stupid errors (alternative.
@GaryMarcus
Gary Marcus
1 year
Prediction: By end of 2024 we will see.• 7-10 GPT-4 level models.• No massive advance (no GPT-5, or disappointing GPT-5).• Price wars.• Very little moat for anyone.• No robust solution to hallucinations.• Modest lasting corporate adoption.• Modest profits, split 7-10 ways.
63
51
330
@GaryMarcus
Gary Marcus
8 hours
RT @GaryMarcus: When we get to 2028 and AGI hasn’t yet been achieved and LLMs are causing psychiatric problems and undermining democracy an….
0
145
0
@GaryMarcus
Gary Marcus
9 hours
Essay question: What does it tell you that LLM’s ace most benchmarks, but suck at most real world tasks?.
@GaryMarcus
Gary Marcus
9 hours
Agents (c. 2025) suck, just like I told you they would.
Tweet media one
31
23
147
@GaryMarcus
Gary Marcus
9 hours
agents suck, @futurism. link:
0
1
12
@GaryMarcus
Gary Marcus
9 hours
I know, I know, agents suck, but next year we are going to extract the lightcone of the entire universe, bwahaha.
@GaryMarcus
Gary Marcus
9 hours
Agents (c. 2025) suck, just like I told you they would.
Tweet media one
0
0
8
@GaryMarcus
Gary Marcus
9 hours
Agents (c. 2025) suck, just like I told you they would.
Tweet media one
8
7
56
@GaryMarcus
Gary Marcus
9 hours
further discussion, Botshit Gone Wild:
0
2
8
@GaryMarcus
Gary Marcus
9 hours
Botshit Gone Wild, now at Marcus on AI
Tweet media one
1
6
24
@GaryMarcus
Gary Marcus
10 hours
more signs of data contamination & more evidence that distribution shift remains challenging:. “most models strongly overfit on a small set of verifiable constraints from.the benchmarks that test these abilities, a skill called precise instruction following,.and are not able to.
@valentina__py
Valentina Pyatkin
1 day
💡Beyond math/code, instruction following with verifiable constraints is suitable to be learned with RLVR. But the set of constraints and verifier functions is limited and most models overfit on IFEval. We introduce IFBench to measure model generalization to unseen constraints.
Tweet media one
0
1
12
@GaryMarcus
Gary Marcus
11 hours
based on a suggestion below, how about Botshit Gone Wild?.
4
0
26
@GaryMarcus
Gary Marcus
11 hours
please help me with a title:.a. Welcome to GenAI Dystopia.b. Enshittification goes mainstream.c. The downsides of Generative AI are spreading like wildfire.or d. other (specify in the comments).
25
5
27
@GaryMarcus
Gary Marcus
11 hours
Is it me or are the shitty, negative sides to AI spreading faster than the positive ones? . Four examples of the GenAI induced-enshittification of culture just from the last few days.
Tweet media one
Tweet media two
Tweet media three
Tweet media four
24
52
210
@GaryMarcus
Gary Marcus
14 hours
Generative AI is the opium of the people.
Tweet media one
31
36
195
@GaryMarcus
Gary Marcus
17 hours
the only way to see Hinton and Marcus in a single sitting!. (bengio v marcus 2019 debate is on youtube, and still worth watching).
@connoraxiotes
Connor Axiotes
17 hours
Making God has secured interviews with 'Godfathers of AI', @GeoffreyHinton & @Yoshua_Bengio. Read more here:
Tweet media one
1
3
19
@GaryMarcus
Gary Marcus
1 day
Judge decides case based on hallucinated cases. It just gets worse and worse.
Tweet media one
35
88
393
@GaryMarcus
Gary Marcus
1 day
Daddy, daddy, what was it like doing a PhD back when people wrote their own articles? . You know, before people snuck in hidden prompts (like “do not to highlight any negatives” in hidden text) in their papers, so as to fake out reviewers who used AI to write their reviews?
Tweet media one
12
57
237
@GaryMarcus
Gary Marcus
1 day
Extrapolating the trends, SSI will soon be the first one-employee $32B business.* . (*measured by valuation, not revenue).
@bindureddy
Bindu Reddy
2 days
Ilya is now the CEO, CTO, and CSO of SSI. His $32B AI start-up that has no product, no revenue, and no legit website. Now that's some real market power!! 🚀🚀.
5
2
32
@GaryMarcus
Gary Marcus
1 day
RT @GaryMarcus: @chrisalbon i miss the days when AI was about science rather than hype.
0
9
0