Vijay

1.9K posts

Vijay

Vijay

@v_kethana

Computer Science @ucberkeley • Research @berkeley_ai @berkeleynlp • Amateur Sanskritist • https://t.co/g3SGopYRiC

Katılım Ekim 2022
370 Takip Edilen289 Takipçiler
Vijay
Vijay@v_kethana·
@clamepending I believe harnesses have lagged behind model capabilities for some time now. So developing good harnesses is very important!
English
0
0
2
13
Mark
Mark@clamepending·
@v_kethana yup! This was just to get an api to buy anything. The thing Im plugging this into is the general prupose agent ui for everyone
English
1
0
1
14
Vijay
Vijay@v_kethana·
@nntaleb & ppl in machine learning too
English
0
0
1
39
Vijay
Vijay@v_kethana·
here is how it would work: - contestants are posed with a challenging open problem in math - they have 48 hours to produce some attempt at a solution - ai tools, computers, internet use are all permitted
English
2
0
3
75
Vijay
Vijay@v_kethana·
imagine if we had hackathons for math
English
1
0
7
106
Vijay
Vijay@v_kethana·
@Satyaki_R Nit: some varieties of recitation do have [ɸ] (surfaces when Visarga + प): e.g. "अन्तःपुरम्" or "अग्निः पूर्वेभिः" (Rigveda 1.1.2). This can be thought of as an "f" sound. Though it is merely an (archaic) positional variant of /p/
English
1
0
1
65
Vijay
Vijay@v_kethana·
@clamepending could this be the "hello world" of robotics?
English
1
0
1
35
Mark
Mark@clamepending·
Bought my own robot arm to see if the cs185 wizardry actually works, haha
English
3
0
2
72
Vijay
Vijay@v_kethana·
@cobbaltt I believe the Twenty-Five Tales of Vetāla is an excellent source of examples of what might be called Medieval (or Later) Sanskrit
English
0
0
2
29
Gopalakrishnan R
Gopalakrishnan R@cobbaltt·
What's "पिता आगच्छन् अस्ति"? Is this a neo-Sanskrit attempt to translate the present perfect in "father has come"? I don't think this is even the periphrastic perfect. Correct me if I'm wrong.
अखिल@akhilesh_41

@ChanchalKamini Complete one🤭👍

English
10
0
21
2.9K
Vijay
Vijay@v_kethana·
@cobbaltt ... i.e. "pitāgacchannasti" is clearly a calque from Hindi (or English). Perhaps surprisingly, Google Gemini grasps this distinction correctly:
Vijay tweet media
English
1
0
0
37
Vijay
Vijay@v_kethana·
@cobbaltt @kmadathil afaik classical Sanskrit has no present progressive in the way Hindi/English does; the closest available form is what you wrote. pitāgacchannasti (=pitā+āgacchan+asti) means something like "father, who is coming, is". while not ungrammatical, it is not quite right
English
1
0
1
45
Vijay retweetledi
Haider.
Haider.@haider1·
Google Jeff Dean says bigger context windows alone are not enough What matters is staged retrieval: lightweight mechanisms that narrow a trillion tokens down to 10 million, then to the million you actually need "you don't need a trillion at once, you need the right million"
English
44
92
982
140.3K
Vijay
Vijay@v_kethana·
I am not saying these tools have SOTA security guarantees; far from it. but the simple design decision of making users approve potentially dangerous commands by default has probably saved us from more headache than we admit.
English
0
0
2
87
Vijay
Vijay@v_kethana·
Dozens of crises have probably been averted by the fact that Codex / Claude Code / Cursor by default ask for permission before making potentially dangerous changes
English
2
0
3
237
Vijay
Vijay@v_kethana·
There's an ongoing undercurrent within the field of RL for language models research: we are underestimating base model capabilities. "Reasoning with Sampling" showed a similar claim to this paper - that base model + the right sampling technique >= RL (for certain tasks)
Yulu Gan@yule_gan

Simply adding Gaussian noise to LLMs (one step—no iterations, no learning rate, no gradients) and ensembling them can achieve performance comparable to or even better than standard GRPO/PPO on math reasoning, coding, writing, and chemistry tasks. We call this algorithm RandOpt. To verify that this is not limited to specific models, we tested it on Qwen, Llama, OLMo3, and VLMs. What's behind this? We find that in the Gaussian search neighborhood around pretrained LLMs, diverse task experts are densely distributed — a regime we term Neural Thickets. Paper: arxiv.org/pdf/2603.12228 Code: github.com/sunrainyg/Rand… Website: thickets.mit.edu

English
0
0
1
167
Vijay
Vijay@v_kethana·
I believe he really meant that English was "a far finer language" for *his* use case.
English
0
0
2
53
Vijay
Vijay@v_kethana·
When Borges said, "I find English a far finer language than Spanish", we should take this with a grain of salt. He admired the language from a distance, growing up in Argentina with a few English family members; so naturally, he viewed English with rose-colored glasses.
English
1
0
2
68
Vijay
Vijay@v_kethana·
The clip of Borges talking about English goes viral every few months; and every time, people misunderstand his point. It's not that English can express things which Spanish can't, but rather that English had certain characteristics which he found preferable to Spanish.
Ryan Rhodes ⚙️🧠@langofmind

Interviewer: Why do you prefer English to Spanish? Me: Well, you see, Spanish is a verb-framed language where the path of motion is encoded in verbs, whereas English is verb-framed, with manner of motion bundled into verb meanings, and that's just groovy, you know

English
1
0
5
185

Keşfet

@clamepending @nntaleb @Satyaki_R @cobbaltt @kmadathil @Konkanist @Saatvata @wagarrwal