(((ل()(ل() 'yoav))))👾 @yoavgo

cs.biu.ac.il/~yogo/ Israel Joined May 2009

Tweets

137K
Followers

65K
Following

2K
Likes

72K

(((ل()(ل() 'yoav))))👾 @yoavgo

7 hours ago

can you explain "serverless backend" to me? it seems to me that their main/only purpose is to hide API keys that you don't want to expose to the client. am I missing something?

5 0 4 2K 1

(((ل()(ل() 'yoav))))👾 @yoavgo

5 days ago

interesting. i initially thought it was going to be a "no DL! how come it has no modern DL??" kinda gripe, but no. its not that at all.

Valeriy M., PhD, MBA, CQF @predict_addict

6 days ago

interesting. i initially thought it was going to be a "no DL! how come it has no modern DL??" kinda gripe, but no. its not that at all.

36 65 1K 208K 1K

Download Image

9 1 16 7K 9

(((ل()(ل() 'yoav))))👾 @yoavgo

5 days ago

amateurs. a proper implementation would have doubled the list when scrolling near the ends.

sky @skydotcs

6 days ago

amateurs. a proper implementation would have doubled the list when scrolling near the ends.

2K 6K 214K 40.7M 16K

Download Video

2 1 18 4K 0

(((ل()(ل() 'yoav))))👾 @yoavgo

6 days ago

i created this thingy yesterday and now I cannot stop watching it. yoavg.github.io/eternal/#

11 4 104 9K 16

arxiv took their time with this paper because apparently they moved it from AI to HCI. it indeed belongs in both, but I still believe it is much more about AI than it is about HCI. but now we have an opportunity to post it again, and it IS a really neat result, imo. so, a win.

Mosh Levy @mosh_levy

a week ago

2 3 16 5K 6

1 1 11 3K 2

Mosh Levy @mosh_levy

a week ago

When reading AI reasoning text (aka CoT), we form a narrative about the underlying computation process, which we take as a transparent explanation of model behavior. But what if our narrative is wrong? We measure that and find it usually is. Now on arXiv: arxiv.org/abs/2508.16599

Mosh Levy @mosh_levy

3 weeks ago

8 27 136 27K 98

Download Image

2 3 16 5K 6

(((ل()(ل() 'yoav))))👾 @yoavgo

a week ago

agents hold a huge promise. building strong, good, interactive and multipurpose ones turns out to be very hard (happy to discuss the challenges). in this project, we attempt to do that, and push forward both useful functionality *and* the science of agent building.

Ai2 @allen_ai

2 weeks ago

9 48 217 293K 97

Download Video

2 5 82 7K 32

Jonathan Blow @Jonathan_Blow

4 weeks ago

Everybody posting about AI in programming sounds like this to me. x.com/LinkedInLunat1…

LinkedIn Lunatics @LinkedInLunat1c

4 weeks ago

Everybody posting about AI in programming sounds like this to me. x.com/LinkedInLunat1…

289 706 19K 956K 1K

Download Image

59 116 2K 89K 164

Mustafa Suleyman @mustafasuleyman

3 weeks ago

@connoraxiotes The debate about whether AI is actually conscious is, for now at least, a distraction. It will seem conscious and that illusion is what’ll matter in the near term

20 4 82 20K 7

(((ل()(ل() 'yoav))))👾 @yoavgo

2 weeks ago

crazy that this is considered controversial / surprising in some circles

Kylie Robison @kyliebytes

2 weeks ago

crazy that this is considered controversial / surprising in some circles

27 37 387 77K 252

Download Image

9 1 63 7K 16

Ai2 @allen_ai

3 weeks ago

LLMs power research, decision‑making, and exploration—but most benchmarks don’t test how well they stitch together evidence across dozens (or hundreds) of sources. Meet MoNaCo, our new eval for question-answering cross‑source reasoning. 👇

10 40 222 20K 98

Download Image

Tomer Wolfson @TomerWolfson

3 weeks ago

Many factual QA benchmarks have become saturated, yet factuality still poses a very real issue! ✨We present MoNaCo, an Ai2 benchmark of human-written time-consuming questions that, on average, require 43.3 documents per question!✨ 📣Blogpost: allenai.org/blog/monaco 🧵(1/5)

1 14 39 4K 12

Download Image

(((ل()(ل() 'yoav))))👾 @yoavgo

3 weeks ago

this is really cool to me, as Reasoning LLMs use language *effectively*. The language-use improves the computation's accuracy. But the work shows they use it in a very different way from us. Its a different way of using the same language!

Mosh Levy @mosh_levy

3 weeks ago

1 0 8 5K 1

1 0 28 4K 10

Mosh Levy @mosh_levy

3 weeks ago

The test is very simple: given a reasoning text with a highlighted target step t, identify the one out of four candidate steps that, if removed, will change the target step. We administered it to 80 participants online. You can test yourself here: do-you-understand-ai.com

1 3 10 2K 1

(((ل()(ل() 'yoav))))👾 @yoavgo

3 weeks ago

in this work, we show that reasoning texts are really not transparent, in the sense that we (humans) are really bad in inferring how the steps (really) depend on each other for the model. yoo can test yourself here: do-you-understand-ai.com

Mosh Levy @mosh_levy

3 weeks ago

8 27 136 27K 98

Download Image

2 14 97 11K 60

Uri Katz @UrikaUri

3 weeks ago

Even when AI reasoning traces are “transparent”, the narrative they encode can differ from the one humans perceive. The transparency debate isn’t just about revealing the model’s steps, it’s about the mismatch between the model’s story and the one we construct from it.

Mosh Levy @mosh_levy

3 weeks ago

8 27 136 27K 98

Download Image

0 2 12 2K 2

Mosh Levy @mosh_levy

3 weeks ago

Producing reasoning texts boosts the capabilities of AI models, but do we humans correctly understand these texts? Our latest research suggests that we do not. This highlights a new angle on the "Are they transparent?" debate: they might be, but we misinterpret them. 🧵

8 27 136 27K 98

Download Image

(((ل()(ل() 'yoav))))👾 @yoavgo

3 weeks ago

welcome back Yanai! what a wonderful addition to the department

Yanai Elazar @yanaiela

3 weeks ago

welcome back Yanai! what a wonderful addition to the department

110 21 523 35K 13

Download Image

1 1 57 5K 0

(((ل()(ل() 'yoav))))👾 @yoavgo

3 weeks ago

did you ever got a paper "stuck" in the arxiv publication queue? how did you resolve it?

4 0 5 3K 1

(((ل()(ل() 'yoav))))👾 @yoavgo

4 weeks ago

if i read things correctly, OpenAI realized that you can get by with much fewer layers if you compensate by generating more tokens ("thinking"). this is huge as it offloads a lot of the cost to the consumer: each gen step is now cheaper to produce, and they sell more of them.