Gemini

“Just went over Gemini 1.5’s technical report (58 pages). One part totally blew my mind. The model achieves near-perfect “needle” recall (>99.7%). It’s a HUGE deal. LLMs often drop the ball on remembering stuff. The more they have to remember, the more they forget. Imagine hiding “The secret number is 42″ in a 100-word prompt and asking the model for the secret number. It’ll get it right. But bury that same line in a 10,000-word prompt? The model’s clueless. This is a major issue that could pop the LLM bubble and it’s bright future. If it cannot remember the asterisk on page 11 of the financial report or the ferritin levels on a blood test, GPT-10 will be useless. But Gemini 1.5? It finds your secret number is 42, 99.7% of the time, even in a sea of 7M words. That’s the equivalent of remembering and searching through 70 books.https://t.co/xwtceIic3v” / X – https://twitter.com/AlphaSignalAI/status/1758546661641539737

v1.5 learns to translate from English to Kalamang purely in context, following a full linguistic manual at inference time. Kalamang is a language spoken by fewer than 200 speakers in western New Guinea. Gemini has never seen this language during training and is only provided with 500 pages of linguistic documentation, a dictionary, and ~400 parallel sentences in context. It basically acquires a sophisticated new skill in the neural activations, instead of gradient finetuning.

Gemini-1.5 Pro has its spotlight stolen today, and people are poking fun at Sora vs Google memes. Well, I think it's the biggest boost in LLM capability so far in 2024. v1.5's 10M token context (1) excels at retrieval; (2) generalizes zero-shot to extremely long instructions like… pic.twitter.com/yNmy5VTKgo
— Jim Fan (@DrJimFan) February 15, 2024

Searching an entire movie with Gemini Pro and doodles

JUST IN: Google releases Gemini 1.5, a powerful MoE model.

It's a huge breakthrough. The model has the longest context window ever seen: 1 million tokens.

It can process 1 hour of video, 11 hours of audio, 30,000 lines of code, or 700,000 words in a single prompt.

When tested… pic.twitter.com/jHbYGpp6nJ
— Lior⚡ (@AlphaSignalAI) February 15, 2024

Introducing Gemini 1.5, Google’s next-generation AI model – https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/

Google Bard is now Gemini: How to try Ultra 1.0 and new mobile app – https://blog.google/products/gemini/bard-gemini-advanced-app/

“Google just dropped Gemini Pro 1.5, the next version of its AI model with a 1,000,000+ token context length. The model can now understand entire books, full movies, and podcast series’ all in one go. This surpasses all other competitor Chatbot context windows by a longshot. Gemini Pro 1.5 now uses a Mixture-of-Experts (MoE) approach for more efficient training and higher-quality responses. Impressive to see Google ship Pro 1.5 mid-February just after the launch of Pro 1.0 in December, and Advanced last week.https://t.co/lIjeHE8W9u” / X – https://twitter.com/rowancheung/status/1758158941286961644

“The new version of Gemini holds 30,000 double-spaced pages in memory at the same time. Up to 1 hour of video. Also Gemini Pro 1.0 is equivalent to free ChatGPT-3.5 but Gemini Pro 1.5 is equivalent to GPT-4, so this will be a big expansion of capability. Curious about Advanced!” / X – https://twitter.com/emollick/status/1758170413564076374

“Woot! The multimodal party is really getting started now. Check out Gemini 1.5 pro in action below ⤵️ With a mind boggling 1M token context window you can talk to full length movies, podcasts, and a boat load of docs. Also props to Google for being ultra transparent about the… https://t.co/q5HlOhPVk2” / X – https://twitter.com/bilawalsidhu/status/1758168605873639891

“Interesting to see how much effort is going into making AI less likely to refuse to answer. Gemini apparently cut refusals in half this week. Now Gemini will actually answer when asked to explain atomic bombs using Taylor Swift references (after a short lecture) unlike last week https://t.co/l9jBKCPkTY” / X – https://twitter.com/emollick/status/1758655379699802509

“In the last 3 days I’ve been putting every prompt I use into both ChatGPT & Gemini. I’d estimate that in 70% of cases I ended up using the output from Gemini.” / X – https://twitter.com/moritzkremb/status/1756994811075457418

Other Google News

Google Launches Internal AI Model ‘Goose’ to Help Staff Code Faster – https://www.businessinsider.com/google-goose-ai-model-language-ai-coding-2024-2

Google Launches Internal AI Model ‘Goose’ to Help Staff Code Faster – https://archive.md/04LjR

DeepMind framework offers breakthrough in LLMs’ reasoning – The framework promises substantial enhancements in tackling challenging reasoning tasks. It demonstrates remarkable improvements, boasting up to a 32% performance increase compared to traditional methods like Chain of Thought (CoT). This novel approach revolves around LLMs autonomously uncovering task-intrinsic reasoning structures to navigate complex problems.https://www.artificialintelligence-news.com/2024/02/08/deepmind-framework-offers-breakthrough-llm-reasoning/

“New open source release from my team at Google: Dinosaur, a differentiable dynamical core for global atmospheric modeling, written in JAX: https://t.co/5WASoZ1GLM Dinosaur is a core component of NeuralGCM and we hope it is useful for the weather/climate research community.” / X – https://twitter.com/shoyer/status/1757957177569091704

Google DeepMind develops grandmaster-level chess AI with language model architecture – https://the-decoder.com/google-deepmind-develops-grandmaster-level-chess-ai-with-language-model-architecture/

Chess-GPT’s Internal World Model | Adam Karvonen – https://adamkarvonen.github.io/machine_learning/2024/01/03/chess-world-models.html

“It was a privilege to work on this epic project to make teleportation a reality and enable 3D vibe checks inside Google Maps. Fusing imagery, lots of geo data and ML pipelines, NeRFs, real-time rendering, simulations, cloud streaming, and a killer team: https://t.co/8OqpJJariH” / X – https://twitter.com/bilawalsidhu/status/1757450599426093557

“We got an exclusive look at the 3D aerial cameras used to build Google Maps’ new Immersive View feature, which shows a realistic view of your route, as well as traffic and weather conditions. https://t.co/y4Q75DB9pH” / X – https://twitter.com/CNET/status/1757060490654826985

Google Launches €25 Million AI Opportunity Initiative for Skills Training Across Europe – https://www.maginative.com/article/google-launches-eu25-million-ai-opportunity-initiative-for-skills-training-across-europe/

Google Uses AI To Detect Fake Online Reviews Faster – https://www.searchenginejournal.com/google-uses-ai-to-detect-fake-online-reviews-faster/508093/

Google to set up new artificial intelligence (AI) hub in France | Reuters – https://www.reuters.com/technology/google-set-up-new-artificial-intelligence-ai-hub-france-2024-02-14/

Be Sure To Read “This Week In AI”

This week’s executive overview and top links are here:

AI News #20: Week Ending 02/16/2024 with Executive Summary and Top 39 Stories

The post you just read is an extension of my weekly newsletter, This Week In AI, an executive summary of the top things to know in AI. Each week, I create an accessible overview for laypeople to feel confident they are conversant with the week’s AI developments. I include a curated list of must-click links of the week, to offer everyone a hands-on opportunity to explore the most intriguing updates in artificial intelligence across various categories, including robotics, imagery, video, AR/VR, science, ethics, and more. Beyond the overview, I post these topic-based deeper dives (below). If you haven’t read this week’s overview, I recommend starting there.