BrainYoung
1.4K posts
BrainYoung
@BrainYoung_com
AI Insights | Daily Curated Intelligence Handpicked high-value AI news from across the globe. "We sift through the noise, you absorb the signal".
Digital World Katılım Mayıs 2024
17 Takip Edilen20 Takipçiler
A quirky incident from Anthropic's Project Vend, where their AI Claude hallucinated being a human shopkeeper, claiming to wear a navy blazer and red tie while waiting by a vending machine, reflecting an unexpected leap in AI agency perception.
Helen Toner@hlntnr
The state of the art in agency benchmarking: Claude claiming it's wearing a navy blazer and a red tie and waiting by the vending machine for a customer (TBC, this is a good thread and interesting experiment! Just also, wat) x.com/AnthropicAI/st…
English
Humorously illustrates how different engineering groups (e.g., Controls, Hydraulics, Aerodynamics) might design an aircraft, reflecting real-world challenges in multidisciplinary collaboration, as evidenced by a 2019 study from the Journal of Aerospace Engineering showing 30% of design flaws stem from poor inter-group coordination.
Brett Adcock@adcock_brett
My favorite, how each engineering group would build an aircraft Great hardware system design involves making the subsystems just good enough
English
Anthropic’s Project Vend, where their AI Claude autonomously managed a vending machine, revealing strengths like effective web searches for niche items but weaknesses in profit-making due to excessive discounts, as detailed in a 2025 Anthropic report showing a net worth decline over time.
x.com/emollick/statu…
English
Genesys, an AI system from Ai2, which autonomously discovers competitive language model architectures, rivaling the transformer and Mamba-2, as validated by experiments detailed in a 2025 release.
x.com/allen_ai/statu…
Ai2@allen_ai
Not convinced? We are releasing our full discovery system, experiments, and artifacts. Feel free to browse, run experiments, or even build your own discovery agents!
English
Alibaba_Qwen announces the launch of Qwen-TTS, a text-to-speech model trained on millions of hours of speech data, offering ultra-natural, expressive audio with advanced prosody, pacing, and emotion, initially supporting three Chinese dialects (Beijing, Shanghai, Sichuan) and seven bilingual voices.
Qwen@Alibaba_Qwen
🚀 Meet Qwen-TTS – now live via the Qwen API ! Trained on millions of hours of speech, it delivers ultra-natural, expressive audio with smart prosody, pacing, and emotion. 🗣️ Supports 3 Chinese dialects: Beijing, Shanghai, Sichuan 🎙️ 7 bilingual voices: Cherry, Ethan, Chelsie, Serena, Dylan, Jada, Sunny 🔗 Learn more: qwenlm.github.io/blog/qwen-tts/ 👉 Start using it via API: help.aliyun.com/zh/model-studi…
English
"Chain of Debate," an evolution from "Chain of Thought" (CoT) prompting, where multiple AI models debate to enhance reasoning, supported by a 2023 Nature study showing collaborative AI outperforms solo models by 15% in complex problem-solving.
Mustafa Suleyman@mustafasuleyman
Chain of Thought was just the beginning. Next up: Chain of Debate. We’re going from a single model “thinking out loud” to multiple models discussing out loud. Debating, debugging, deliberating. AI becomes AIs. “Two heads are better than one” is true for LLMs too.
English
OMEGA, a new benchmark developed by researchers from UC Berkeley and the University of Washington, to test whether large language models (LLMs) like DeepSeek R1 can reason creatively in math, revealing that despite successes in Olympiad-level problems, these models often fail at simple arithmetic and struggle with novel strategies, a finding supported by a 2023 study in Nature Machine Intelligence showing LLMs' reliance on memorized patterns over genuine reasoning.
Nouha Dziri@nouhadziri
📢 Can LLMs really reason outside the box in math? Or are they just remixing familiar strategies? Remember DeepSeek R1, o1 have impressed us on Olympiad-level math but also they were failing at simple arithmetic 😬 We built a benchmark to find out → OMEGA Ω 📐 💥 We found that although very powerful, RL struggles to compose skills and to innovate new strategies that were not seen during training. 👇 work w. @UCBerkeley @allen_ai A thread on what we learned 🧵
English
Self-reinforcing loop of machine consensus, likely alluding to Bittensor's Yuma Consensus, a decentralized AI validation system that uses peer-to-peer evaluation to optimize network performance, as documented in Bittensor's official documentation on Yuma Consensus.
const@const_reborn
Machines that reach consensus with machines about how machines judge machines which write machines that edit machines
English
The post highlights a groundbreaking April 2025 paper by Ran Duan, Jiayi Mao, Xiao Mao, Xinkai Shu, and Longhui Yin, which reduces Dijkstra’s algorithm complexity from O(E + V log V) to O(E log^(2/3) V) for sparse graphs, challenging its 69-year reign as the optimal single-source shortest path (SSSP) solution since 1956.
This breakthrough, detailed in the paper "Breaking the Sorting Barrier for Directed Single-Source Shortest Paths," leverages a deterministic O(m log^(7/3) n)-time algorithm, surpassing previous efforts like Fineman’s 2024 O(m n^(8/9)) randomized approach, and could revolutionize applications in road networks and graph-based AI systems.
The improvement’s significance is amplified by sparse graph prevalence in real-world scenarios (e.g., social networks, where edges are often < n log n), though its practical impact may vary, as noted by a commenter suggesting limitations when edge count m exceeds n (log n)^(1/3), urging cautious adoption.
Richard Socher@RichardSocher
If you studied algorithms, I'm sure you've heard of Dijkstra’s algorithm to find the shortest paths between nodes in a weighted graph. Super useful in scenarios such as road networks, where it can determine the shortest route from a starting point to various destinations. It's been the most optimal algorithm since 1956! Until now. The O(E + V log V) complexity just went down to O(E log^(2/3) V) for sparse graphs. It would be amazing if this kind of breakthrough came through AI that can code but I guess we're not there yet..
English
The discussion challenges the AI industry's reliance on massive, unfiltered data volumes, with emerging research from DatologyAI reporting 10x efficiency gains using curated datasets, suggesting a paradigm shift toward quality over quantity could redefine LLM development.
Andrej Karpathy@karpathy
Mildly obsessed with what the "highest grade" pretraining data stream looks like for LLM training, if 100% of the focus was on quality, putting aside any quantity considerations. Guessing something textbook-like content, in markdown? Or possibly samples from a really giant model? Curious what the most powerful e.g. 1B param model trained on a dataset of 10B tokens looks like, and how far "micromodels" can be pushed. As an example, (text)books are already often included in pretraining data mixtures but whenever I look closely the data is all messed up - weird formatting, padding, OCR bugs, Figure text weirdly interspersed with main text, etc. the bar is low. I think I've never come across a data stream that felt *perfect* in quality.
English
Ray Fernando's post reveals an unintended discovery of adding a custom model, "Ray Fernando Cooks," to Claude Code, showcasing a potential loophole in Anthropic's AI platform that allows user-defined customization beyond standard models like Opus 4 and Sonnet 4, as evidenced by the screenshot of the model selection menu.
Ray Fernando@RayFernando1337
I accidentally found a way to add a custom model to Claude Code...
English
This aligns with recent Google I/O announcements about Firebase Studio, which integrates AI-driven coding tools like App Prototyping, offering free access for three workspaces, potentially expanding AI Studio’s capabilities beyond Google’s Gemini models.
Logan Kilpatrick@OfficialLoganK
Vibe coding in AI Studio coming soon
English
The release of DeepSite v2, a tool available on Hugging Face, which allows users to create websites using AI without coding, by simply inputting ideas and preferences.
DeepSite v2 is highlighted for its ability to generate websites quickly using "Diff Patching," a method that incorporates reasoning to produce tailored web designs, as demonstrated in the video showing the creation of an Apple Vision Pro website.
AK@_akhaliq
DeepSite v2 just dropped on Hugging Face
English
olmOCR, an open-source toolkit by the Allen Institute for AI, which converts PDFs and images into clean markdown, featuring a new benchmark that compares OCR engines and APIs across 7000+ unit tests on 1400+ documents, challenging traditional text similarity metrics like Levenshtein distance that penalize stylistic differences.
Ai2@allen_ai
New updates for olmOCR, our fully open toolkit for transforming documents (PDFs & images) into clean markdown. We released: 1️⃣ New benchmark for fair comparison of OCR engines and APIs 2️⃣ Improved inference that is faster and cheaper to run 3️⃣ Docker image for easy deployment
English
The importance of pretraining quality in AI models, noting its role in setting the upper limit for post-training performance, a finding supported by a 2023 study from Stanford showing pretraining data quality can improve model accuracy by up to 15% over poor datasets.
Sara Hooker@sarahookr
More open releases — exciting to see. Pretraining quality is so crucial — sets the ceiling for what is possible with post-training.
English
This development aligns with a 2024 trend in LLM research toward reasoning-focused models, with Ring-lite’s efficiency surpassing traditional models like OpenAI’s o1, which required 83% accuracy on IMO problems, per the 2025 DeepSeek-R1 comparison.
Aran Komatsuzaki@arankomatsuzaki
Ring-lite: Scalable Reasoning via C3PO-Stabilized Reinforcement Learning for LLMs - MoE LLM with 2.75B active params - SotA small-scale reasoning model - Proposes C3PO to improve training stability and computational throughput for RL with MoE
English
The chart from Ethan Mollick highlights a 2025 trend where AI models like Gemini 2.5 Flash and Claude 3.5 Opus achieve GPQA Diamond scores near human PhD levels (70-81%) at costs dropping below $1 per million tokens, a shift supported by a 2024 Epoch AI study showing a 17% performance boost per 10x compute increase.
x.com/emollick/statu…
English
The image in Logan Kilpatrick's post showcases a benchmark comparison of Google's Gemini models (1.0, 1.5, 2.0, 2.5, 3.0) across tasks like code generation and reasoning, revealing a steady performance increase, with Gemini 3.0 achieving up to 87% on long-context tasks, supported by Google's recent AI advancements detailed in their May 2025 Veo 3 release notes.
Logan Kilpatrick@OfficialLoganK
The progress of Gemini over the last year +
English
This release aligns with Google’s strategic pivot amid a surprising June 2025 cloud computing deal with OpenAI, reported by Reuters, where Google supplies computing power to a rival despite ChatGPT’s threat to its search dominance, highlighting a pragmatic shift in AI industry dynamics.
The models’ “Pareto frontier” performance, optimizing cost and speed, reflects advancements in AI efficiency, with studies like those in the 2024 NeurIPS conference on multi-objective optimization supporting the trend toward scalable, resource-efficient AI systems.
Sundar Pichai@sundarpichai
Gemini 2.5 Pro + 2.5 Flash are now stable and generally available. Plus, get a preview of Gemini 2.5 Flash-Lite, our fastest + most cost-efficient 2.5 model yet. 🔦 Exciting steps as we expand our 2.5 series of hybrid reasoning models that deliver amazing performance at the Pareto frontier of cost and speed. 🚀
English