All about AI, Web 3.0, BCI
3.81K subscribers
777 photos
29 videos
162 files
3.55K links
This channel about AI, Web 3.0 and brain computer interface(BCI)

owner @Aniaslanyan
Download Telegram
RoboCasa is a large-scale simulation framework of everyday tasks

Researchers use generative AI tools to create diverse objects, scenes, and tasks. Simulation plays a pivotal role in our Data Pyramid for training generalist robots.

It's all open-source of course.
Big news โ€“ Gemini 1.5 Flash, Pro and Advanced results are out!๐Ÿ”ฅ

- Gemini 1.5 Pro/Advanced at #2, closing in on GPT-4o
- Gemini 1.5 Flash at #9, outperforming Llama-3-70b and nearly reaching GPT-4-0125 (!)

Pro is significantly stronger than its April version. Flashโ€™s cost, capabilities, and unmatched context length make it a market game-changer!

In Chinese, Gemini 1.5 Pro & Advanced are now the best #1 model in the world. Flash becomes even stronger!

Gemini family remains top in our new "Hard Prompts" category, which features more challenging, problem-solving user queries.

Learn more about Hard Prompts.
Researchers introduced SignLLM.

It's the first multilingual Sign Language Production (SLP) AI model capable of generating avatar videos of sign language gestures from prompts across eight languages.
What is going on? OpenAI has said its mission is not to build "superintelligence" in an apparent backtrack from previous comments by Sam Altman, as it readies its new model.
Mistral released first code model, Codestral-22B:
- Outperforms all open code models
- Fluent in 80+ languages, 32K context length
- Available on our API and for free on Le Chat
- Integrated with VS Code

Weights.
Scale launched SEAL Leaderboardsโ€”private, expert evaluations of leading frontier models.

Evaluations are a critical component of the AI ecosystem.

Evals are incentives for researchers, and our evaluations set the goals for how we aim to improve our models.

Trusted 3rd party evals are a missing part of the whole ecosystem, which is why Scale built these.

They eval'd many of the leading models:

- GPT-4o
- GPT-4 Turbo
- Claude 3 Opus
- Gemini 1.5 Pro
- Gemini 1.5 Flash
- Llama3
- Mistral Large

On Coding, Math, Instruction Following, and Multilinguality (Spanish).
โค4
Researchers introduced Geometry-Informed Neural Networks to train shape generative models
without any data (!!)
, combining learning under constraints, neural fields as a suitable representation, and generating diverse solutions to under-determined problems.

Paper.
AI_AMA_1717076968.pdf
1.3 MB
๐—ฅ๐—ฒ๐—ฝ๐—ผ๐—ฟ๐˜: ๐—ง๐—ต๐—ฒ ๐—™๐˜‚๐˜๐˜‚๐—ฟ๐—ฒ ๐—ผ๐—ณ ๐—›๐—ฒ๐—ฎ๐—น๐˜๐—ต โ€“ ๐—ง๐—ต๐—ฒ ๐—˜๐—บ๐—ฒ๐—ฟ๐—ด๐—ถ๐—ป๐—ด ๐—Ÿ๐—ฎ๐—ป๐—ฑ๐˜€๐—ฐ๐—ฎ๐—ฝ๐—ฒ ๐—ผ๐—ณ ๐—”๐˜‚๐—ด๐—บ๐—ฒ๐—ป๐˜๐—ฒ๐—ฑ ๐—œ๐—ป๐˜๐—ฒ๐—น๐—น๐—ถ๐—ด๐—ฒ๐—ป๐—ฐ๐—ฒ ๐—ถ๐—ป ๐—›๐—ฒ๐—ฎ๐—น๐˜๐—ต ๐—–๐—ฎ๐—ฟ๐—ฒ

๐—ž๐—ฒ๐˜† ๐—›๐—ถ๐—ด๐—ต๐—น๐—ถ๐—ด๐—ต๐˜๐˜€:

1. ๐—›๐—ถ๐˜€๐˜๐—ผ๐—ฟ๐—ถ๐—ฐ๐—ฎ๐—น ๐—–๐—ผ๐—ป๐˜๐—ฒ๐˜…๐˜: AI has been a part of medicine since the mid-20th century, with its role significantly expanding in fields like radiology, cardiology, and neurology.
2. ๐—ง๐—ฒ๐—ฐ๐—ต๐—ป๐—ผ๐—น๐—ผ๐—ด๐—ถ๐—ฐ๐—ฎ๐—น ๐—”๐—ฑ๐˜ƒ๐—ฎ๐—ป๐—ฐ๐—ฒ๐—บ๐—ฒ๐—ป๐˜๐˜€: Recent innovations in deep learning and foundation models are unlocking new use cases in healthcare, from cancer prognoses to predicting adverse clinical events.
3. ๐—ฃ๐—ต๐˜†๐˜€๐—ถ๐—ฐ๐—ถ๐—ฎ๐—ป ๐—ฃ๐—ฒ๐—ฟ๐˜€๐—ฝ๐—ฒ๐—ฐ๐˜๐—ถ๐˜ƒ๐—ฒ๐˜€: While 65% of surveyed physicians see definite or some advantage in using AI, 70% express concerns about potential biases, privacy risks, and liability issues.
4. ๐—ฆ๐—ฝ๐—ฒ๐—ฐ๐—ถ๐—ฎ๐—น๐˜๐˜†-๐—ฆ๐—ฝ๐—ฒ๐—ฐ๐—ถ๐—ณ๐—ถ๐—ฐ ๐—จ๐˜€๐—ฒ ๐—–๐—ฎ๐˜€๐—ฒ๐˜€: Each medical specialty leverages AI differently, from emergency medicine's vitals monitoring to family medicine's focus on personalized patient education and medication adherence.
5. ๐—”๐—ฑ๐—บ๐—ถ๐—ป๐—ถ๐˜€๐˜๐—ฟ๐—ฎ๐˜๐—ถ๐˜ƒ๐—ฒ ๐—ฆ๐˜‚๐—ฝ๐—ฝ๐—ผ๐—ฟ๐˜: 56% of surveyed physicians identify administrative burden reduction through automation as AI's biggest opportunity.
Astronomers are preparing to use AI to tackle 300 petabytes of data annually.

Cecilia Garraffo's AstroAI initiative is pioneering the fusion of AI and astronomy to explore deep cosmic questions, already planning dozens of projects with a 50-member interdisciplinary team.
The Simulationโ€™s new Showrunner platform lets you direct, star in, and even get paid for your own AI-generated TV shows

The lines are blurring fast between creators and audiences โ€” and the traditional Hollywood media model is changing in front of our eyes.
Similarity is Not All You Need: Endowing Retrieval-Augmented Generation with Multiโ€“layered Thoughts.
Microsoft presents Self-Exploring Language Models: Active Preference Elicitation for Online Alignment

SELM significantly boosts the performance on instructionfollowing benchmarks such as MT-Bench and AlpacaEval 2.0

Repo.
Google DeepMind released new demos of its Veo AI video generation model.

The demo showcases the ability to turn single reference images into new videos with simple text instructions.
Revolutionary work from Huggingface: FineWeb

FineWeb:
0. Pretraining is far less intuitive than instruct finetune
1. Unclear what data to include to boost performance
2. HF's team simply tried different filters [3] -> trained small models on the filtered dataset -> measured eval scores.

Result:
The first version of FineWeb. A pre-training dataset that is was methodically selected for the exact purpose of making the models you train on it get the best eval scores as possible.

FineWeb-Edu:
0. Filtered the dataset even more to include only "high educational value" texts. [1]
2. Used Llama-3-70B-Instruct [2] to extract "educational score" for 500K texts.
3. Trained a classifier to classify the rest of the 15T tokens.

Result:
FineWeb-Edu outperform every other open pre-training dataset by a huge margin.

Other treasures in the blog.
In the blog there is a ton of useful material: how to filter at scale? how to run LLMs at scale? breakdown of every single filter contribution to the score..


To the best this is the first large scale empirical study of pre-training data trying to improve an underlying model's performance.

[1] As proposed on "Textsbooks are all you need" (phi-1): arxiv.org/abs/2306.11644

[2] Language filter -> Gopher filter (from it's paper) -> MinHash -> C4 filter (simple rules in the paper) -> PII filter.

[3] Prompt from arxiv.org/abs/2401.10020
Real_Estate_Toksenization_1717419791.pdf
3 MB
Real estate tokenisation will unlock $3.5 trillion in global liquidity by 2027. This could solve one of the biggest issues in real estate โ€“
Illiquidity.


This report explores the detailed process of tokenising real estate, the legal and regulatory challenges, and what this could mean for the industry.

Here are key takeaways:

1. Tokenisation offers many benefits, such as increased liquidity, fractional ownership, and streamlined transactions.

2. The tokenisation process involves multiple phases, including deal structuring, digitisation, primary distribution, post-tokenisation management, and secondary trading.

3. Regulations for tokenised securities differ by country, with some countries being more open to these innovations than others.

4. Determining the value and tax obligations of tokenised assets is complex. We need to tackle these issues very carefully.

5. Challenges also include confidentiality and legal uncertainties as the legal frameworks are still evolving.

6. Opportunities extend beyond traditional investments to things like employee incentives, lease-to-own models, and co-working spaces.
Fascinating finding! Research suggests that the #brain language regions are hardly activated when "reading" python and other programming languages

Instead, reading computer code appears to engage both left and right sides of the Multiple Demand Network (MDN).

The MDN, whose activity is spread throughout the frontal and parietal lobes of the brain, is typically recruited for tasks that require holding many pieces of information in mind at once, and is responsible for our ability to perform a wide variety of mental tasks.

Cognitive skills, such as reading, writing, map-based navigation, mathematical reasoning, and scientific logic, draw heavily on the MDN.

โ†–The left hemisphere is activated more when solving math and logic problems .
โ†— The right hemisphere activates more when doing tasks involving spatial navigation.

In a companion paper appearing in the same issue of eLife, a team of researchers from Johns Hopkins University also reported that solving code problems activates the multiple demand network rather than the language regions.

โ€œThese findings suggest there isnโ€™t a definitive answer to whether coding should be taught as a math-based skill or a language-based skill. In part, thatโ€™s because learning to program may draw on both language and multiple demand systems, even if โ€” once learned โ€” programming doesnโ€™t rely on the language regions,โ€ the researchers say.
๐Ÿฆ„3
So it begins, Hollywood studios being open about their leaning into AI strategies.

A cost reducing measure, going to be interesting to see the impacts of this.
๐Ÿ’…3
The Bank for International Settlements (BIS) is launching Project Rialto to explore how instant cross-border payments could be improved using a modular foreign exchange component combined with settlement in wholesale central bank digital currencies ( #wCBDC ).