Links for 2024-08-02

AI:

1. Google released an experimental updated version of Gemini 1.5 Pro that is #1 on the Chatbot Arena. Try it here: https://aistudio.google.com/app/

2. Method prevents an AI model from being overconfident about wrong answers https://news.mit.edu/2024/thermometer-prevents-ai-model-overconfidence-about-wrong-answers-0731

3. Sparse Autoencoders as a microscope for AI internals. https://deepmind.google/discover/blog/gemma-scope-helping-the-safety-community-shed-light-on-the-inner-workings-of-language-models/

4. Diffusion Augmented Agents: A Framework for Efficient Exploration and Transfer Learning https://arxiv.org/abs/2407.20798

5. Odyssey equips LLM-agents with advanced skills for exploring Minecraft. https://github.com/zju-vipa/Odyssey?tab=readme-ov-file

6. Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge https://arxiv.org/abs/2407.19594

7. “We let models make hundreds or thousands of attempts when solving a problem, rather than just one...outperforming the single-attempt SOTA.” https://scalyresearch.stanford.edu/pubs/large_language_monkeys/

8. “Which is better, running a 70B model once, or a 7B model 10 times? Our findings reveal that the repeated use of smaller models can yield consistent improvements.” https://arxiv.org/abs/2404.00725

9. Achieving new SOTA standards by ensembling every other model into a meta-model that learns when to call each LLM. https://www.notdiamond.ai/

10. Claude Engineer https://github.com/Doriandarko/claude-engineer

11. LangGraph Studio: The first agent IDE https://www.youtube.com/watch?v=pLPJoFvq4_M

12. “By making programs differentiable, we inherently introduce probability distributions over their execution, providing a means to quantify the uncertainty associated with program outputs.” https://arxiv.org/abs/2403.14606

13. How AI is changing warfare https://www.economist.com/briefing/2024/06/20/how-ai-is-changing-warfare [no paywall: https://archive.is/yw8Yz]

14. “We discover a systematic way to scale up robot data...and we multiply that data 1000x or more in simulation.” https://x.com/DrJimFan/status/1818302152982343983

15. Figure AI: "Only recently has time opened a window of opportunity to scale billions of intelligent humanoid robots…Life is about to turn into a SciFi film." https://x.com/adcock_brett/status/1819191267785581049

Health:

1. One dose of a new nasal spray treatment clears toxic tau proteins from brain cells, improving memory. https://www.utmb.edu/news/article/utmb-news/2024/07/03/new-breakthrough-in-alzheimer-s-research--utmb-researchers-develop-nasal-spray-treatment-for-alzheimer-s-disease

2. New weight-loss drugs are causing people to spend less on groceries and choose healthier options. A new study shows that users buy 52% less snacks and confectionery, 47% less baked goods, and 28% less sugary drinks. https://nypost.com/2024/07/27/lifestyle/weight-loss-drugs-eat-into-grocery-basket/

Physics:

1. Is nature really as strange as quantum theory says? Neutron measurements prove: It doesn't work without the strange properties of quantum theory. https://www.tuwien.at/en/phy/ati/news/neutronen-auf-klassisch-unerklaerlichen-bahnen-1

2. New work suggests that when black holes die, they turn into white holes. And that these objects are an ideal candidate for the dark matter that cosmologists believe fills the universe but have never directly observed. https://arxiv.org/abs/2407.09584

Miscellaneous:

1. Space is a latent sequence: A theory of the hippocampus https://www.science.org/doi/10.1126/sciadv.adm8470

2. Probability is just...really weird https://www.youtube.com/watch?v=zczGnnM05TQ

3. How computers work explained from scratch. https://www.youtube.com/playlist?list=PLnAxReCloSeTJc8ZGogzjtCtXl_eE6yzA

4. List of biotech founders and drug hunters who were unlikely to succeed (and yet they did) https://www.ladanuzhna.xyz/writing/list-of-biotech-founders

5. Romae Industriae: What were the binding constraints on a Roman Industrial Revolution? https://www.maximum-progress.com/p/romae-industriae

"You can tell just by looking at this progression that this was a labor of love for a lot of very smart people." – Paul Graham

Links for 2024-08-04

AI:

1. AgentGen uses LLMs to synthesize diverse environments and planning tasks in a scalable way. https://arxiv.org/abs/2408.00764

2. Using LLM embeddings to capture word-by-word linguistic content transmitted from the speaker's brain to the listener's brain in real-time, face-to-face conversations https://www.cell.com/neuron/fulltext/S0896-6273(24)00460-4

3. An introduction to reinforcement learning for neuroscience https://arxiv.org/abs/2311.07315

4. From Text to Life: On the Reciprocal Relationship between Artificial Life and Large Language Models https://arxiv.org/abs/2407.09502

5. Toward De Novo Protein Design from Natural Language https://www.biorxiv.org/content/10.1101/2024.08.01.606258v1

6. The newly released Palmyra-Fin-70B outperforms Claude 3.5 Sonnet, GPT-4o, and Mixtral-8x7b on the long-fin-eval benchmark, across a variety of real-world financial use cases. https://x.com/rohanpaul_ai/status/1819443015481446643

7. “TPU transformation: A look back at 10 years of our AI-specialized chips” https://cloud.google.com/transform/ai-specialized-chips-tpu-history-gen-ai

8. Tyler Cowen on ChatGPT Advanced Voice Mode: "It’s happening, and this is to date one of the most vivid and impressive illustrations of what is possible. A mere three years ago this would have seemed like witchcraft." https://marginalrevolution.com/marginalrevolution/2024/08/chatgpt-advanced-voice-mode.html

9. ChatGPT Advanced Voice Mode Impresses Testers With Sound Effects, Catching Its Breath https://arstechnica.com/information-technology/2024/07/when-counting-quickly-openais-new-voice-mode-stops-to-catch-its-breath/

10. “I'm not going to make any arguments about what the future holds. I just want to provide a list of 50 conversations that I (a programmer and research scientist studying machine learning) have had with different large language models to meaningfully improve my ability to perform research and help me work on random coding side projects.” https://nicholas.carlini.com/writing/2024/how-i-use-ai.html

11. Character.AI CEO Noam Shazeer returns to Google. Google is also signing a non-exclusive agreement with Character.AI to use its tech. https://techcrunch.com/2024/08/02/character-ai-ceo-noam-shazeer-returns-to-google/

12. UK government shelves £1.3bn tech and AI plans https://www.bbc.com/news/articles/cyx5x44vnyeo

Miscellaneous:

1. “From an evolutionary perspective, what distinguishes the human brain? You may say, the neocortex. Surprisingly, in humans and other great apes, the expansion of the cerebellum accelerated faster than the enlargement of the cerebral cortex.” https://www.cell.com/current-biology/fulltext/S0960-9822(14)01069-0

2. “No matter what you post on social media. You can be found. Whether it's a zoomed in photo of your table or just a photo of your lunch. Even the smallest details in a photo give the biggest hints.” https://www.youtube.com/watch?app=desktop&v=Ue94gpWqEkM

Politics:

1. Iran has told Arab diplomats that they don't care if the response triggers a war with Israel, according to people familiar with the conversations https://www.wsj.com/world/middle-east/iran-rebuffs-calls-for-restraint-in-its-response-to-killing-of-hamas-leader-309314e7 [no paywall: https://archive.is/tmWJ3]

2. “Why is society so vulnerable to far-left ideas? One key reason: it’s hard to counter weaponized empathy. When actions are taken under the banner of a long-suffering group, that makes it much more difficult to challenge the worldview behind them. And far-left activists do skew female, suggesting that empathy plays a significant role.” https://x.com/RichardMCNgo/status/1819400985569329350

3. "In 2017, a survey of economists by the Chicago Booth School of Business asked if refugees will benefit Germany. Only 6% said they would be net cost. Almost 1 million Syrians in Germany, over half on welfare, the rest get medical, housing benefits. Not great forecasting." https://x.com/whyvert/status/1819395080714871003

Predictions for the future of software engineering: https://x.com/russelljkaplan/status/1820460524460802256

Links for 2024-08-06

AI:

1. Figure 02 unveiled — working autonomously at BMW's Spartanburg factory. ⦿ New 16 Degrees of Freedom hand ⦿ Onboard inference, running VLM locally for speech-to-speech reasoning ⦿ 2.25 KWh battery ⦿ Exoskeleton structure + Integrated wiring https://www.youtube.com/watch?v=0SRVJaOg9Co (press article: https://spectrum.ieee.org/figure-new-humanoid-robot)

2. Fully-automatic robot dentist performs world's first human procedure https://newatlas.com/health-wellbeing/robot-dentist-world-first/

3. Big tech’s huge AI spending isn’t slowing down. And according to their forward-looking statements, that spending is expected to go up even more. https://sherwood.news/tech/meta-amazon-microsoft-massive-ai-capex-spending-quarterly-earnings/

4. Flux: OpenAI’s DALL-E 3-Like AI For Free, Forever! https://www.youtube.com/watch?v=-7crpGKEA2g

5. Meta presents Self-Taught Evaluators: Without any labeled preference data, the proposed model utperforms commonly used LLM judges such as GPT-4 and matches the performance of the top-performing reward models trained with labeled examples https://arxiv.org/abs/2408.02666

6. AI capabilities can be significantly improved without expensive retraining https://arxiv.org/abs/2312.07413

7. MiniCPM-V: A GPT-4V Level MLLM on Your Phone https://arxiv.org/abs/2408.01800

8. DeepL’s latest large language model, which is trained to specialize in translation, outperforms Google Translate and GPT-4 for translation tasks. https://thenextweb.com/news/deepl-new-llm-that-outperforms-google-translate-chatgpt

9. A New Type of Neural Network Is More Interpretable -- Kolmogorov-Arnold Networks could point physicists to new hypotheses https://spectrum.ieee.org/kan-neural-network

10. GPUDrive: Data-driven, multi-agent driving simulation at 1 million FPS https://arxiv.org/abs/2408.01584

11. Tenstorrent has developed a new set of AI chips that are much less expensive than NVIDIA’s. They are available as PCIe cards or as components of complete workstations. https://wccftech.com/tenstorrent-wormhole-ai-processors-risc-v-phenomenal-price-to-performance-value/

12. Meta says it will need 10x more computing power to train Llama 4 compared to Llama 3. https://techcrunch.com/2024/08/01/zuckerberg-says-meta-will-need-10x-more-computing-power-to-train-llama-4-than-llama-3/

13. “Breaking my hand forced me to write all my code with AI for 2 months. I’m never going back.” https://erikschluntz.com/software/2024/07/30/code-with-ai.html

Miscellaneous:

1. ‘Sensational’ Proof Delivers New Insights Into Prime Numbers https://www.quantamagazine.org/sensational-proof-delivers-new-insights-into-prime-numbers-20240715/

2. Neuroscience research into people with aphantasia, who don’t experience mental imagery, is revealing how imagination works and demonstrating the sweeping variety in our subjective experiences. https://www.quantamagazine.org/what-happens-in-a-mind-that-cant-see-mental-images-20240801/

3. No proof that radiation from X rays and CT scans causes cancer https://www.sciencedaily.com/releases/2016/02/160203134456.htm

4. Japan's unmanned stores count on shoppers' honesty https://web-japan.org/trends/11_tech-life/tec202309_unmanned-stores.html

5. "In my opinion, every moment beyond this should simply be interpreted within the context of a lower moment. Every even moment (4 - aka kurtosis, 6, 8, etc.) corresponds to variance, while every odd moment (5, 7, 9, etc.) corresponds to skewness. As the moments get larger, they are more impacted by outliers. So, the fourth moment (kurtosis) measures the same things that the second moment does (variance), but with a heavier focus on the outliers. This is where the "fat tails" description of kurtosis comes from. It measures the spread of the data but is more depend on the behavior of outliers in the tails." https://www.reddit.com/r/AskStatistics/comments/6d3fsp/comment/di0b0pc/

Empirical data on how useful AI agents are currently compared to humans: They can't do everything, but they can do a decent chunk of what humans can do, and they can do it significantly cheaper/faster.

Read more: https://metr.org/blog/2024-08-06-update-on-evaluations/

[Open Source] Unitree First View Teleoperation for Humanoid Robots to advance the convenience of data collection for humanoid robots: https://github.com/unitreerobotics/avp_teleoperate

Links for 2024-08-08

AI:

1. “Can LLMs predict results of social science experiments? Across 70 studies, we find striking alignment (r = .85) between simulated and observed effects. Overall our results show high accuracy of LLM-derived predictions for experiments with human participants, generally greater accuracy than samples of lay and expert humans.” https://docsend.com/view/qeeccuggec56k9hd

2. “LLaVA-OneVision allows strong transfer learning across different modalities/scenarios, yielding new emerging capabilities. In particular, strong video understanding and cross-scenario capabilities are demonstrated through task transfer from images to videos.” https://arxiv.org/abs/2408.03326

3. Key-Point-Driven Mathematical Reasoning Distillation of Large Language Model https://arxiv.org/abs/2407.10167

4. Benchmarking LLMs for Optimization Modeling and Enhancing Reasoning via Reverse Socratic Synthesis https://arxiv.org/abs/2407.09887

5. "Transformers are Universal In-context Learners": in this paper, we show that deep transformers with a fixed embedding dimension are universal approximators for an arbitrarily large number of tokens. https://arxiv.org/abs/2408.01367

6. “How can we prevent LLM safeguards from being simply removed with a few steps of fine-tuning? We show it's surprisingly possible to make progress on creating safeguards that are tamper-resistant, reducing malicious use risks of open-weight models.” https://arxiv.org/abs/2408.00761

7. Diffusion Models as Data Mining Tools https://arxiv.org/abs/2408.02752

8. Hierarchical Conditioning of Diffusion Models Using Tree-of-Life for Studying Species Evolution https://arxiv.org/abs/2408.00160

9. Google announces Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters — Test-time compute can be used to outperform a 14× larger model https://arxiv.org/abs/2408.03314

10. A New Study Says AI Models Encode Language Like the Human Brain Does https://singularityhub.com/2024/08/07/a-new-study-says-ai-models-encode-language-like-the-human-brain-does/

11. A.I. ‐ Humanity's Final Invention? https://www.youtube.com/watch?v=fa8k8IQ1_X0

12. AI “godfather” Yoshua Bengio has joined a UK project to prevent AI catastrophes https://www.technologyreview.com/2024/08/07/1095879/ai-godfather-yoshua-bengio-joins-uk-project-to-prevent-ai-catastrophes/ [no paywall: https://archive.is/wcpgo]

Miscellaneous:

1. “We're using ultrasound to safely and non-invasively measure and modulate brain activity at high resolution” https://quintinfrerichs.xyz/nudge

2. Japanese scientists develop simplified EUV scanner that can make production of chips considerably cheaper https://www.tomshardware.com/tech-industry/japanese-scientists-develop-simplified-euv-scanner-that-can-make-production-of-chips-considerably-cheaper

3. Tiny arm bone belonged to smallest ancient human ever found https://www.nature.com/articles/d41586-024-02548-6

4. “The implications for life in the liquid water oceans, under the surface of icy moons, are obvious, and enormous. So I'm going to predict now, with medium confidence (and a couple of caveats, to follow) that we may well ultimately discover similar polymetallic nodules, producing oxygen through similar chemical processes, on the warm seafloors of the liquid water oceans under the frozen crusts of icy moons.” https://theeggandtherock.com/p/the-deep-ocean-floor-is-covered-in

5. Feasibility of keeping Mars warm with nanoparticles https://www.science.org/doi/10.1126/sciadv.adn4650

6. “When that enormous magnitude-9 earthquake hit Japan in 2011, it caused waves 1.5 meters high in some lakes in NORWAY!” https://mathstodon.xyz/@johncarlosbaez/112920894947197795

Politics:

1. ‘Sky’s the limit’: Fort Stewart soldiers prepare for the modern battlefield by building small drones from scratch https://www.stripes.com/branches/army/2024-08-06/army-soldiers-building-drones-fort-stewart-14761022.html

2. What can we say about the "far right" riots? https://www.aporiamagazine.com/p/what-can-we-say-about-the-far-right

Google unveils "Achieving Human Level Competitive Robot Table Tennis"! The robot won 100% vs. beginners and 55% vs. intermediate players, showcasing solid amateur human-level performance.

"The robot has to be good at low level skills, such as returning the ball, as well as high level skills, like strategizing and long-term planning to achieve a goal.

The robot first trains in a simulated environment, which can model the physics of table tennis matches accurately.

Once deployed to the real world, it collects data on its performance against humans to refine its skills back in simulation - creating a continuous feedback loop."

Read more: https://sites.google.com/view/competitive-robot-table-tennis/home

Links for 2024-08-09

AI:

1. Chinese open weights model easily surpasses all previous models, both closed and open, at MATH https://qwenlm.github.io/blog/qwen2-math/

2. Using LLMs to close the expertise gap of humans, empowering general non-expert human programmers to match experienced competitive programmers (including IOI medalists). https://arxiv.org/abs/2406.04604

3. Optimus-1: Hybrid Multimodal Memory Empowered Agents Excel in Long-Horizon Tasks https://cybertronagent.github.io/Optimus-1.github.io/

4. Agent K builds itself in order to complete tasks for you. Its mind is a bunch of agents that collaborate to complete tasks. Those agents will collaborate to develop new agents if they're needed to complete a given task. https://github.com/mikekelly/AgentK

5. Transformer Explainer: Interactive Learning of Text-Generative Models https://arxiv.org/abs/2408.04619

6. CodexGraph: Bridging Large Language Models and Code Repositories via Code Graph Databases https://arxiv.org/abs/2408.03910

7. Terence Tao’s recent lecture on AI https://www.youtube.com/watch?v=_sTDSO74D8Q

8. LG unleashes South Korea's first open-source AI, challenging global tech giants https://venturebeat.com/ai/lg-unleashes-south-koreas-first-open-source-ai-challenging-global-tech-giants/

9. Equivariant neural networks and piecewise linear representation theory https://arxiv.org/abs/2408.00949

Miscellaneous:

1. Efficient coding with chaotic neural networks: A journey from neuroscience to physics and back https://arxiv.org/abs/2408.01949

2. “The novel 3D printing method uses sound waves, instead of light or heat, to create solid material out of a polymer solution from behind a physical barrier.” https://engineering.ucdavis.edu/news/uc-davis-researchers-win-manufacturing-award-vision-3d-print-inside-human-body

3. Your microwave oven has its own microbiome https://www.nature.com/articles/d41586-024-02553-9 [archived version: https://archive.is/ZV7gO]

4. Chinese megaconstellation launch creates field of space debris https://spacenews.com/chinese-megaconstellation-launch-creates-field-of-space-debris/

As Russia's “3-day special operation” stretches into its 900th day, let's examine Vladimir Putin's unintended accomplishments:

1. Brought the war to his own soil
2. Expanded NATO by two historically neutral nations (Finland and Sweden)
3. Caused a revival of defense spending in the West
4. Renewed Western appreciation for their military forces
5. Boosted Western arms exports
6. Bootstrapped Western autonomous weapons technology
7. Turned Russia into a totally dependent Chinese vassal state
8. Lost most of his soft power over Western politicians
9. Tarnished Russia's superpower image.

At the same time, NATO hasn't lost a single square meter of territory or a single soldier, while Russia has lost more than 4,534 officers, according to Russian sources.

In light of these outcomes, Putin's “special operation” can only be described as a strategic catastrophe for Russia.

Sakana AI announces The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

“Our system is capable of executing the entire ML research lifecycle: from inventing research ideas and experiments, writing code, to executing experiments on GPUs and gathering results.

The AI Scientist can produce entire scientific papers that exceed the acceptance threshold at a top machine learning conference as judged by our automated reviewer.

In one run the agent tried to change its own code by removing some obstacles, to better achieve its (completely unrelated) goal.”

Read more: https://sakana.ai/ai-scientist/
Code: https://github.com/SakanaAI/AI-Scientist

Links for 2024-08-13

AI:

1. Introducing Genie... the most capable AI software engineering system. It achieves state-of-the-art on SWE-Bench with 30.08%. That's a 57% improvement! https://cosine.sh/blog/genie-technical-report

2. Open and closed-ended problem solving in humans and AI: The influence of question asking complexity https://www.sciencedirect.com/science/article/pii/S1871187124001366

3. rStar: a self-play mutual reasoning approach that significantly improves reasoning capabilities of small language models (SLMs) without fine-tuning or superior models. https://arxiv.org/abs/2408.06195

4. Tree Attention, an exact attention approach with less communication and memory requirements than Ring Attention, enabling more efficient scaling to million token sequence lengths https://arxiv.org/abs/2408.04093

5. Combining GraphRAG and VectorRAG leads to a HybridRAG system that outperforms both individually. https://arxiv.org/abs/2408.04948

6. Transformers are energy-based models in disguise. Tree Attention: Topology-aware Decoding for Long-Context Attention on GPU clusters https://arxiv.org/abs/2408.04093

7. The ChatGPT of databases. 100% open source. https://postgres.new/

8. China uses LLaMa-3 to train a semiconductor advice LLM https://arxiv.org/abs/2408.00804

9. A clip from the GPT-4o safety card where the voice model suddenly yells "No!" and then starts imitating the user's voice https://www.reddit.com/r/singularity/comments/1enne2l/comment/lh7zsb4/

10. “One thing I was very wrong about ~4yrs ago is how fundamental “synthetic data” in ML would be.” https://x.com/PreetumNakkiran/status/1821928149908848869

11. OpenAI expert Scott Aaronson on consciousness, quantum physics and AI safety https://scottaaronson.blog/?p=8200

12. Paul Graham: "I was just talking with a friend who's been investing in startups for about 20 years, and we both agreed that one of the weirdest things about the AI boom is seeing journalists writing their usual contrarian stories about how it's a bubble about to burst...But this thing is real. If anything, alarmingly real." https://x.com/paulg/status/1823125140944945568

Biotech:

1. Is It Ethical To Hand-Pick Your Child’s Genes? https://www.youtube.com/watch?v=e3cXRs60xiU

2. Why Does Ozempic Cure All Diseases? https://www.astralcodexten.com/p/why-does-ozempic-cure-all-diseases

3. A bacterial antiviral immune machinery creates a new gene de novo just from an RNA to fend off viruses. https://www.biorxiv.org/content/10.1101/2024.05.08.593200v1

4. A Novel Treatment Slashes HIV Up To 10,000-Fold in Monkeys With Just a Single Dose https://singularityhub.com/2024/08/12/a-novel-treatment-slashes-hiv-up-to-10000-fold-in-monkeys-with-just-a-single-dose/

5. "When the results of a new drug that prevented 100% of HIV cases were announced at the 2024 AIDS conference, the room burst into spontaneous Applause" https://blogs.jwatch.org/hiv-id-observations/index.php/lenacapavir-prep-trial-brings-down-the-house-at-the-international-aids-conference/2024/07/25/

Neuroscience:

1. Demonstration that sublinear summation in dendrites can unlock the computation of nonlinear functions by a single neuron https://www.nature.com/articles/s41598-024-65866-9

2. “By combining photochemical sectioning with volumetric lattice light-sheet imaging and petabyte-scale computation, we imaged and reconstructed axons and myelination sheaths across entire mouse olfactory bulbs at nanoscale resolution.” https://www.biorxiv.org/content/10.1101/2024.08.01.605857v1

Miscellaneous:

1. “The single most undervalued fact of linear algebra: Matrices are graphs, and graphs are matrices. Encoding matrices as graphs is a cheat code, making complex behavior simple to study.” https://x.com/svpino/status/1822966303642308903

2. Billions of dollars of venture capital is flowing into defense-tech startups focused on futuristic, AI-enabled weapons. Palmer Luckey’s Anduril is their biggest bet. https://www.wsj.com/tech/anduril-drones-palmer-luckey-china-ukraine-china-951494ec [no paywall: https://archive.is/uWfOR]

Agent Q - bringing next-generation AI agents with planning and AI self-healing capabilities, with a 340% improvement over LLama 3's baseline zero-shot performance!

Not only does their fine-tuned LLaMa 70B outperform GPT4 - it goes from 18.6%-81.7% zero-shot performance after a single day of autonomous self-play! If they allow for online search absolute success rate jumps up to 95.4%!

Read more: https://www.multion.ai/blog/introducing-agent-q-research-breakthrough-for-the-next-generation-of-ai-agents-with-planning-and-self-healing-capabilities