Links for 2023-10-21
Can GPT-4 teach a robot hand to do pen spinning tricks better than you do? Eureka, an open-ended agent that designs reward functions for robot dexterity at super-human level.
Eureka exploits the remarkable zero-shot generation, code-writing, and in-context improvement capabilities of state-of-the-art LLMs, such as GPT-4, to perform evolutionary optimization over reward code. The resulting rewards can then be used to acquire complex skills via reinforcement learning. Without any task-specific prompting or pre-defined reward templates, Eureka generates reward functions that outperform expert human-engineered rewards.
…using Eureka rewards in a curriculum learning setting, we demonstrate for the first time, a simulated Shadow Hand capable of performing pen spinning tricks, adeptly manipulating a pen in circles at rapid speed.
Project page: https://eureka-research.github.io/
Paper: https://arxiv.org/abs/2310.12931
Thread: https://twitter.com/DrJimFan/status/1715397393842401440
“We are excited to announce CLIN 🤖: The first continually learning language agent that excels in both task adaptation and generalization to unseen tasks and environments in a pure zero-shot setup.” https://allenai.github.io/clin/
Llemma: open LMs for math trained on up to 200B tokens of mathematical text. The performance of Llemma 34B approaches Google's Minerva 62B despite having half the parameters. https://arxiv.org/abs/2310.10631 (Models/data/code: https://github.com/EleutherAI/math-lm)
Improving Large Language Model Fine-tuning for Solving Math Problems https://arxiv.org/abs/2310.10047
“Language models are bad a basic math...There are a few reasons why numbers are hard. The main one is Tokenization. When training a tokenizer from scratch, you take a large corpus of text and find the minimal byte-pair encoding for a chosen vocabulary size.” https://twitter.com/andrew_n_carr/status/1714326003030638848
CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules https://arxiv.org/abs/2310.08992
When can transformers reason with abstract symbols? https://arxiv.org/abs/2310.09753
Interactive Task Planning with Language Models https://arxiv.org/abs/2310.10645
“…beliefs about drug intake can be so powerful that it can even influence the outcome in a dose-dependent manner, something that was thought only possible with the actual drugs.” https://twitter.com/doctorveera/status/1713749819376779412
“This fraudulent paper claiming that GMOs cause cancer in rats was used by several countries including Kenya & India to ban or halt GM crops. Sèralini should be asked to pay for damages done to farmers in these countries for being denied a valuable product because of his mischief” https://twitter.com/AgBioWorld/status/1714254643092426845
Ukraine links:
“I wouldn't be surprised if the Donetsk offensive has cost around 20,000+ KIA/WIA on the Russian side since it kicked off. UA estimates are unlikely not inflated either. If anything, they are on the low end.” https://twitter.com/IhateTrenches/status/1715372233147711862
“Railway tracks area near Bakhmut. Russian soldiers scattered around the fields. Footage from Shershen company, part of the 3rd assault brigade.” https://twitter.com/NOELreports/status/1715344831705227277
Avdiivka front. Russian TOS-1 220mm MRLS split into atoms by the FPV drone of the 59th Brigade of Ukraine. 🌞https://twitter.com/bayraktar_1love/status/1715324855527084044
Avdiivka, repelling Russian attack. By the Special Operations Forces of Ukraine. https://twitter.com/bayraktar_1love/status/1715300569689059648
Further meat assaults at Avdiivka https://twitter.com/GloOouD/status/1715367044042899902
“8 helicopters were completely destroyed, 6 were damaged and probably totalled and a radar was destroyed at Berdyansk. A Pantsir SAM located outside the image to the east may also have been damaged.” https://twitter.com/COUPSURE/status/1715357927056273533
“Your $600 donated Wild Hornet drone destroyed a rare ultra-modern Buk-M3 launcher (possibly worth $50M+)” https://twitter.com/ArmedMaidan/status/1715345841571938497
"Measures are being taken to improve the defensive lines on the northern state border. More than 500.000 anti-tank mines have been installed in the main directions of the enemy's probable attack since June 2022," Lieutenant General Serhii Naev said. https://suspilne.media/598745-na-osnovnih-napramkah-imovirnogo-nastupu-rf-na-pivnoci-vstanovleno-ponad-500-tisac-min-naev/
Not sure what’s going on here but it looks like these Russians might be drunk? https://twitter.com/NOELreports/status/1715478412578418966
“We have emails and documents from members of GRU Unit 29155–Putin’s assassination and sabotage squad—proving their culpability for a 2011 bombing in Bulgaria. The IEDs were planted in Czechia.The target was ammunition bound for Georgia.” https://theins.ru/en/politics/266039
Israel links:
“Iran's underground missile and drone base. It can withstand any bunker-busting missile and nuclear strike. The new "Fattah" hypersonic missiles are also kept here.” https://twitter.com/Megatron_ron/status/1715275275800289522 (German analog: https://www.youtube.com/watch?v=0BJJNKzQM20)
“Al Arabiya journalist grilling Hamas’s Khalid Mashaal from an Arab perspective” https://twitter.com/arash_tehran/status/1715354932595847322
“Interesting footage of another al Qassam drone dropped PG-7R/al Yassin tandem HEAT RPG projectile against an Israeli Merkava Mk.4 tank east of Khan Yunis on thursday, causing damage to the rear of the turret.” https://twitter.com/CalibreObscura/status/1715333209557209278
“Now that the Ashli hospital flap is over, the media must prepare for a story that will be difficult to navigate. The Al-Shifa hospital in Gaza City is also the command center for Hamas. At some point, it will be an Israeli target. A legitimate military target, per the laws of war” https://twitter.com/JSchanzer/status/1715340775981240533
French Intel Says No Sign Gaza Hospital Blast Was 'Israeli Strike' https://www.barrons.com/news/french-military-intel-says-no-indication-gaza-hospital-blast-was-israeli-strike-946e673e