Links for 2024-07-21

AI:

1. Interfacing an LLM with a reliable symbolic system (Prolog) raises math performance near ceiling: Tested on an *entirely new* collection of math word problems, the Non-Linear (NLR) reasoning dataset, to ensure all were outside the LLM training set. GPT fails completely. But GPT writing prolog code succeeds near ceiling. https://arxiv.org/abs/2407.11373

2. How can informal reasoning improve formal theorem proving? Lean-STaR: A framework for learning to interleave informal thoughts with steps of formal proving. Training language models to produce informal thoughts prior to each step of a proof, thereby improving the model’s formal theorem-proving capabilities. https://arxiv.org/abs/2407.10040

3. Adding self-modeling to artificial networks causes a significant reduction in network complexity. When artificial networks learn to predict their internal states as an auxiliary task, they change in a fundamental way. https://arxiv.org/abs/2407.10188

4. A system that incorporates both natural language pre-training and reinforcement learning from the start. https://arxiv.org/abs/2308.01399

5. Georgia Tech researchers have developed a neural network, RTNet, that mimics human decision-making processes, including confidence and variability, improving its reliability and accuracy in tasks like digit recognition. https://research.gatech.edu/new-neural-network-makes-decisions-human-would

6. R+X: Retrieval and Execution from Everyday Human Videos — By using a VLM for retrieval and in-context IL for execution, robots can now learn from unlabelled videos of humans performing tasks. https://www.robot-learning.uk/r-plus-x

7. “We trained GPT2 to predict the product of two numbers up to 🌟20🌟 digits w/o intermediate reasoning steps, surpassing our previous 15-digit demo! How does a 12-layer LM solve 20-digit multiplication w/o CoT?🤯” https://arxiv.org/abs/2405.14838

8. AI AI Bias: Large Language Models Favor Their Own Generated Content https://arxiv.org/abs/2407.12856

9. Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive? https://arxiv.org/abs/2406.04391

10. NuminaMath datasets: the largest collection of ~1M math competition problem-solution pairs, ranging in difficulty from junior challenge to Math Olympiad preselection. These datasets were used to win the 1st Progress Prize of the AI Math Olympiad. https://huggingface.co/collections/AI-MO/numinamath-6697df380293bcfdbc1d978c

11. The AI-Powered Future of Coding Is Near https://www.wired.com/ai-powered-coding/ [no paywall: https://archive.is/BRBCw]

https://x.com/TolgaBilge_/status/1814828193985003666

Miscellaneous:

1. Accidentally exposed yellowish-green crystals reveal ‘mind-blowing’ finding on Mars, scientists say https://edition.cnn.com/2024/07/20/science/nasa-curiosity-rover-mars-sulfur-rocks/index.html

2. Chinese nuclear reactor is completely meltdown-proof https://www.newscientist.com/article/2440388-chinese-nuclear-reactor-is-completely-meltdown-proof/ [no paywall: https://archive.is/OPcrZ]

3. Your brain on shrooms — how psilocybin resets neural networks https://www.nature.com/articles/d41586-024-02275-y

4. Dogs might have evolved to read your emotions https://www.nature.com/articles/d41586-024-02320-w

Ukraine:

1. “Ukrainian attack drones struck Millerovo Airfield in Rostov Oblast, a major forward airbase near the Ukrainian border used by Russian strike aircraft. The airbase is currently burning, with multiple massive fires along the flightline.” https://x.com/Osinttechnical/status/1814520339658850769 (“Better satellite images from Millerovo Air Base published by Radio Svoboda show that at least a maintenance hangar and a fuel and lubricants storage facility were destroyed, with the latter partially burned out. The damage was inflicted by Ukrainian drones.” https://x.com/kromark/status/1815026901510946821)

2. Satellite images showing the aftermath of the Ukrainian naval drones attack on a Russian military base in Novoozerne (Donuzlav Lake) in occupied Crimea. https://x.com/kromark/status/1814410503193571628

3. Strike on Russian BUK air defence system with classic air defence missile detonation. https://x.com/bayraktar_1love/status/1814708449638297750

4. UK Government and defence industry stepping up support for Ukraine https://www.gov.uk/government/news/uk-government-and-defence-industry-stepping-up-support-for-ukraine

5. 'Trump said that it was a “very good call” and that he will work to negotiate a peace deal...A person close to Zelenskyy...said it went “exceedingly well” and that Trump pledged to “achieve a just peace in Ukraine”' https://www.politico.com/news/2024/07/19/trump-zelenskyy-ukraine-call-00169905

6. Zelensky: "We can end the hot stage of the war before the end of this year ... It doesn't mean that all territories must be recaptured ... But we must be strong on the battlefield ... we need to take steps to force it to retreat ... this will give (us) better position at the negotiating table." https://x.com/wartranslated/status/1814989166901440555

7. How Ukrainian tanks stay alive in a war of drones https://www.youtube.com/watch?v=HZb0okchAFY

8. “There have been people claiming Russia is about to Run out of equipment since day two of the war, and they were Wrong. However, I come back to the point from the start. Russia's store of equipment is deep but not bottomless, We don't truly know how close to the end we are, but for some significant types at least it might be close.” https://x.com/verekerrichard1/status/1814284243465368063

9. “And here’s what Chasiv Yar looks like from aside.” https://x.com/IAPonomarenko/status/1815030122367316381

10. Hungary facing fuel crisis as Ukraine turns up heat on Russian oil supplies https://www.politico.eu/article/hungary-fuel-crisis-ukraine-sanctions-russian-oil-imports-lukoil-central-europe/