semih duru @semihduru

Joined January 2010

Tweets

38
Followers

20
Following

127
Likes

300

semih duru @semihduru

2 years ago

@OctopusEnergy there are bills from 2023 appeared on my account which i paid already that time. What is happening with your systems? It is impossible now to track and understand them ☹️

1 0 0 36 0

View Details

🚨New paper!🚨 Self-Rewarding LMs - LM itself provides its own rewards on own generations via LLM-as-a-Judge during Iterative DPO - Reward modeling ability improves during training rather than staying fixed ...opens the door to superhuman feedback? arxiv.org/abs/2401.10020 🧵(1/5)

5 212 1K 466K 814

View Details

Hyung Won Chung @hwchung27

3 years ago

I gave a talk at Seoul National University. I titled the talk “Large Language Models (in 2023)”. This was an ambitious attempt to summarize our exploding field. Video: youtu.be/dbo3kNKPaUA Slides: docs.google.com/presentation/d… Trying to summarize the field forced me to think about what really matters in the field. While scaling undeniably stands out, its far-reaching implications are more nuanced. I share my thoughts on scaling from three angles: 1) Change in perspective is necessary because some abilities only emerge at a certain scale. Even if some abilities don’t work with the current generation LLMs, we should not claim that it doesn’t work. Rather, we should think it doesn’t work yet. Once larger models are available many conclusions change. This also means that some conclusions from the past are invalidated and we need to constantly unlearn intuitions built on top of such ideas. 2) From first-principles, scaling up the Transformer amounts to efficiently doing matrix multiplications with many, many machines. I see many researchers in the field of LLM who are not familiar with how scaling is actually done. This section is targeted for technical audiences who want to understand what it means to train large models. 3) I talk about what we should think about for further scaling (think 10000x GPT-4 scale). To me scaling isn’t just doing the same thing with more machines. It entails finding the inductive bias that is the bottleneck in further scaling. I believe that the maximum likelihood objective function is the bottleneck in achieving the scale of 10000x GPT-4 level. Learning the objective function with an expressive neural net is the next paradigm that is a lot more scalable. With the compute cost going down exponentially, scalable methods eventually win. Don’t compete with that. In all of these sections, I strive to describe everything from first-principles. In an extremely fast moving field like LLM, no one can keep up. I believe that understanding the core ideas by deriving from first-principles is the only scalable approach.

41 578 3K 568K 3K

View Details

semih duru @semihduru

6 years ago

@imTolstoyevski Neden olmasin ? youtu.be/g83QsCdP2lM

0 0 1 0 0

View Details

Lex Fridman @lexfridman

6 years ago

GPT-3 has 175 billion parameters (synapses). Human brain has 100+ trillion synapses. In 2020, it costs ~$2.6 billion to train neural net the size of the human brain. If ML training efficiency improvement trends continue, in 2040, it'll cost $80,000. Video: youtube.com/watch?v=kpiY_L…

42 421 2K 0 148

View Details

faraaz @faraaz

6 years ago

I got GPT-3 to start writing my SQL queries for me p.s. these work against my *actual* database!

122 1K 7K 0 508

View Details

hasan söylemez @hasansoylemez

6 years ago

İnsan kendi içine doğru bir yolculuk yapmak istiyorsa fiziksel olarak da bir yolculuk yapmalı. En zoru, insanın kendisiyle yüzleşmesidir. Yalnız ve uzun yolculuklarda bunu yaşamak kaçınılmazdır. Bazen kendinizden nefret edersiniz. Geçmişle, hayatla, insanlarla hesaplaşırsınız.

19 180 2K 0 28

View Details

Andrej Karpathy @karpathy

6 years ago

new blog post: "Biohacking Lite" karpathy.github.io/2020/06/11/bio… fun to write / learn about quick tour of human energy metabolism

58 355 2K 0 478

View Details

Eric Berger @SciGuySpace

6 years ago

Finally, it feels like the 21st century.

288 3K 39K 0 126

View Details

Tim Rocktäschel @_rockt

7 years ago

I am proud to announce the release of the NetHack Learning Environment (NLE)! NetHack is an extremely difficult procedurally-generated grid-world dungeon-crawl game that strikes a great balance between complexity and speed for single-agent reinforcement learning research. 1/

14 178 690 0 85

View Details

Greta Thunberg @GretaThunberg

7 years ago

Hang in there! Help is available.

Mark Humphries @markhumphries

7 years ago

THE GRETA THUNBERG HELPLINE: For adults angry at a child.

4K 80K 206K 0 5K

5K 81K 378K 0 4K

View Details

The Economist @TheEconomist

7 years ago

"Susamam" ("I can't stay quiet") features a parade of 20 rappers calling out Turkey's injustices, crime and corruption econ.st/34CujDz

2 92 290 0 6

View Details

Ece Temelkuran @ETemelkuran

7 years ago

Important note for European countries and the US: Turkey is so far the first country that defeated the new global wave of authoritarianism through democratic means- and after 17 years of such an oppresive regime. Hopefully this experince will be studied by the global opposition.

24 394 2K 0 5

View Details

Fırat Tanış @firattanis

7 years ago

o ceketi çıkarttırmayacaktınız

684 19K 148K 0 54

View Details

Barış Yarkadaş @barisyarkadas

7 years ago

ŞU AN...

297 1K 17K 0 8

View Details

Ece Temelkuran @ETemelkuran

7 years ago

Follow Turkey-Istanbul today to see what it takes to recover a lost democracy. Millions will be defending their right to be political subjects without any expectation from state institutions or any guarantee of security. May the human dignity win today in my city, Istanbul.

11 102 627 0 0

View Details

Linus Tech Tips @LinusTech

7 years ago

Make Mac Pro Grate Again

76 799 5K 0 13

View Details

dgntkn @dgntkn

8 years ago

Binali Dahaİmamoğlu

Arşiv Hesabı @TCMeclisBaskani

8 years ago

İstanbullulara müjdeli haber: Daha önce söz verdiğimiz gibi; kişi sayısı daha fazla olan hanelerde daha çok indirim olmak üzere, su tüketiminde aylık ortalama yüzde 40 indirim yapılacak. Öğrenci Akbil'i aylık 85 liradan 40 liraya düşürülecek. Hayırlı olsun. #DahaGüzelOlacak