Links for 2024-09-14
AI:
Tencent presents GameGen-O: Open-world Video Game Generation https://gamegen-o.github.io/
Llama-Omni: The open-source answer to GPT-4o real-time speech interaction and it is based on Llama-3.1-8B-Instruct. https://github.com/ictnlp/LLaMA-Omni
Transformers solve an open problem in symbolic mathematics: discovering Lyapunov functions https://www.youtube.com/live/yCzV97QNG8w?si=vMchebo08p23Vq31&t=300
Manufacturing aware generative model architectures https://www.jura.bio/blog/variationalsynthesis
“Conspiracy beliefs famously resist correction, ya? WRONG: We show brief convos with GPT4 to reduce conspiracy beliefs by ~20%! Lasts over 2 months. Works on entrenched beliefs. Tailored AI response rebuts specific evidence offered by believers.” https://www.science.org/doi/10.1126/science.adq1814
Estimating Tail Risk in Neural Networks https://www.lesswrong.com/posts/xj5nzResmDZDqLuLo/estimating-tail-risk-in-neural-networks
OpenAI o1:
“This language model now has a strong text-based 'system 2' a la Kahneman. A very small fraction of people are as good at formal reasoning as this model…The next thing I would do is to marry a multimodal version of this with reinforcement-learning training of problem-solving, planning, and agency in a sophisticated simulated environment. There's no obvious obstacle (I know of) to doing this (Deepmind's GATO was a tiny but seemingly-successful test case)…” https://x.com/AnthonyNAguirre/status/1834646562146075077
Terence Tao tested OpenAI o1 (presumably o1-preview): “It is certainly a more capable tool than previous iterations, though still struggling with the most advanced research mathematical tasks.” https://mathstodon.xyz/@tao/113132502735585408
“We put OpenAI o1 to the test against ARC Prize. Results: both o1 models beat GPT-4o. And o1-preview is on par with Claude 3.5 Sonnet. Can chain-of-thought scale to AGI? What explains o1's modest scores on ARC-AGI?” https://arcprize.org/blog/openai-o1-results-arc-prize
“We worked closely with OpenAI over the last few weeks to evaluate OpenAI o1's reasoning capabilities with Devin. We found that the new series of models is a significant improvement for agentic systems that deal with code.” https://www.cognition.ai/blog/evaluating-coding-agents
“This morning I had my first visceral “🤯” moment with AI for ~2 years 🧵on o1 and cryptic crosswords” https://x.com/matthewclifford/status/1834485810113990786
⚡️ o1-mini is now the leading model (or maybe agent?) for both logical and mathematical reasoning. https://huggingface.co/spaces/allenai/ZeroEval
“There is no guarantee the summarizer is faithful, though we intend it to be. I definitely do not recommend assuming that it's faithful to the CoT, or that the CoT itself is faithful to the model's actual reasoning!” https://x.com/polynoamial/status/1834644274417119457
One important piece of information OpenAI gave us is that they have internal access to powerful specialized versions of o1.
Now that some rumors have been confirmed, I think we can be more confident that others will turn out to be partially true. Specifically, OpenAI is using specialized versions of o1 to generate synthetic data for their next big model. In other words, GPT-5 is coming.
Strawberry easily becomes a data flywheel. If the answer is correct, the entire search trace becomes a mini dataset of training examples, which contain both positive and negative rewards.
Neuroscience:
Researchers use propofol to uncover the interactions between the thalamus and cortex that underlie consciousness https://www.michiganmedicine.org/health-lab/how-brains-inner-chamber-governs-your-state-consciousness
Wellesley team’s new research on anesthesia unlocks important clues about the nature of consciousness: “Wiest and his research team found that when they gave rats a drug that binds to microtubules, it took the rats significantly longer to fall unconscious under an anesthetic gas. The research team’s microtubule-binding drug interfered with the anesthetic action, thus supporting the idea that the anesthetic acts on microtubules to cause unconsciousness.” https://www.wellesley.edu/news/wellesley-teams-new-research-on-anesthesia-unlocks-important-clues-about-the-nature-of-consciousness
“Certain spontaneous activities in our brain known as local field potential events (LFPs) were able to give decisive indicators regarding how our brains work. These spontaneous signals seem to play an important role in how our brains process information even in the absence of external stimuli.” https://www.fau.eu/2024/09/09/news/ai-uncovers-the-secrets-of-human-cognition/
Miscellaneous:
Discovery of a new phase of matter in 2D which defies normal statistical mechanics https://www.eurekalert.org/news-releases/1057250
“These results strongly suggest Neanderthal-derived DNA is playing a significant role in autism susceptibility across major populations in the United States.” https://www.nature.com/articles/s41380-024-02593-7
WebGPU Puzzles: Learn GPU Programming in Your Browser https://www.answer.ai/posts/2024-09-12-gpupuzzles.html
HARMONY OF RESILIENCE: Recorded in space and sent to Earth via SpaceX’s Starlink constellation https://x.com/PolarisProgram/status/1834557770374296010
American politics:
“During the debate, Trump sputtered that FBI homicide data are underreported by police departments. In reality the FBI estimates compensate for these missing data, & Datalytics estimates, taken directly from big-city depts, confirm there's no current crime surge. On the contrary, after the 20-21 spike (which was genuine), we've returned to the lowest homicide rate since 1963.” https://x.com/sapinker/status/1834652412830716228
“VIOLENT CRIME IS DECLINING POST-2021 AND IS WAY BELOW HISTORICAL LEVELS” https://x.com/eyeslasho/status/1834272919557403042
Ukraine:
Russians try to destroy a 'dragon drone' that is destroying their positions. https://x.com/NOELreports/status/1834878497157709828
“Russian forces continue counterattacks along the entire perimeter of the Ukrainian bridgehead in Kursk Oblast but have achieved only minor successes due to the continued offensive and counterattacks by the "Siversk" OTG. The enemy regained control of Apanasovka, Byakhovo, Vishniovka, Viktorovka, Vnezapnoe, Gordeevka, Krasnooktyabrskoe, Obukhovka, part of Snagost, and the settlement "10 Let Oktyabrya." The enemy has advanced in those areas of Kursk Oblast that are not fully controlled by "Siversk" OTG, but will face greater challenges when counterattacking in areas where the Ukrainian Defense Forces have already consolidated.” https://cdsdailybrief.substack.com/p/russias-war-on-ukraine-130924
Footage of a new breakthrough by Ukrainian forces in the Kursk region, in the direction of the village of Veseloe https://x.com/Danspiun/status/1834916620759621775
The 🇺🇦Ukrainian Air Assault Force repelled the 🇷🇺Russian mechanised attack in Kursk region https://x.com/Tendar/status/1834946249436356833
“The 🇺🇦Ukrainian 66th Mech Brigade shows the result of unsuccessful attacks on their positions and, as a result bodies of the dead invaders.” https://x.com/GloOouD/status/1834611770410799122
“The Ukrainian Main Directorate of Intelligence (GUR) released a longer video of their operation against the Russian-occupied gas rig “Crimea-2”. The operation was undertaken in the night from September 10 to 11. It shows the gas rig heavily hit, including several explosions.” https://x.com/Tendar/status/1834950562695020751
“This is Europe's worst humanitarian tragedy of the 21st century.” https://x.com/Tatarigami_UA/status/1834675823388840125
Russia's 2022 invasion of Ukraine has been an extinction-level event for the former USSR's arsenal.
Russian visually confirmed losses of armored combat vehicles [tanks, AFVs, IFVs, APCs, and MRAPs] since the start of their 2022 invasion of Ukraine have exceeded 10,000: https://www.oryxspioenkop.com/2022/02/attack-on-europe-documenting-equipment.html
Tanks: 3 371
AFVs: 1,561
IFVs: 4,534
APCs: 489
MRAPs: 56
--------------
Total: 10,011
Large losses of equipment, in general, are also confirmed by satellite imagery, which shows a massive decrease: https://x.com/CovertCabal/status/1834661369678848135
Putin is once again attempting to sway Western politicians by drawing another so-called "red line," suggesting that crossing it could trigger war with NATO.
However, this is unlikely for several reasons:
1. Russia Cannot Win a Conventional War Against Europe: In a conventional conflict, Russia would struggle even against Europe alone, and a nuclear war would be tantamount to suicide. Putin is not willing to commit national suicide over Ukraine. Even if he were irrational enough to consider it, his generals and most oligarchs are not. They are unlikely to sacrifice their own lives and those of their families for a war that Putin initiated and could end at any time.
2. History Shows Nuclear Powers Avoid Using Nukes, Even When Losing Wars: There are multiple precedents where nuclear-armed nations have lost or nearly lost wars without resorting to nuclear weapons. The Soviet Union didn’t use them when it withdrew from Afghanistan, the U.S. didn’t resort to them in Vietnam, and Israel refrained from using them despite facing a multi-front attack from Arab states. In the case of Russia, there is no existential threat justifying the catastrophic diplomatic, economic, and military consequences of using nuclear weapons simply because its military campaign in Ukraine is being resisted.
More importantly: Succumbing to Nuclear Blackmail Increases the Risk of Nuclear Conflict
Caving to nuclear threats only emboldens further aggression. If the use of nuclear threats achieves foreign policy goals, more threats will follow, increasing the chances of miscalculation and escalation. Worse, other countries may conclude that they, too, must acquire nuclear weapons—either to assert their foreign policy aims or to defend themselves—further heightening the risk of nuclear conflict.