Alex Brooker’s Post

On the current LLM war, Google strikes back... sort of. Today Google announced its Gemini 1.5 with 1 Million tokens, among other goodies. The thing is: it is not available yet with such amount of tokens. But current 1.5 version is a good competitor of GPT4.5. It's exciting to see Google stepping up their game in the language model space with the announcement of Gemini 1.5. While the 1 million tokens feature is not yet available, the current version is already proving to be a strong competitor to GPT4.5. It will be interesting to see how other players in the space, such as OpenAI, respond to this development. Let's hope this doesn't turn into a vaporware war. https://lnkd.in/ekPSzi69 #Google #Gemini #LanguageModels #AI #OpenAI

Here is key AI developments in last week: 1) Claude 3.5 Haiku released by Anthropic. It surpasses all previous Claude models (except the new 3.5 Sonnet) on coding and agentic tasks.  - 200k context window  - 50% discount with the Batches API   However the model is significantly expensive at the rate of $1 per million input tokens and $5 per million output tokens then previous rates of $0.25 and $1.25, respectively. This makes seven times more expensive than OptnAI's GTP-4o Mini and more costly than Google's Gemini 1.5 Flash. 2) In a Reddit AMA, OpenAI CEO Sam Altman admitted that it is delaying it's product roadmap due to lack of compute capacity. While Microsoft being significantly investor in OpenAI at the same time competing with it's own AI system Copilot, could it be that Microsoft is not giving sufficient compute to OpenAI to prioritizing for it's own Copilot AI's need?   3) Physical Intelligence released released general-purpose AI model Pie Zero for physical world. It's first version of the model perform various tasks such as folding cloths, making coffee, folding box, pouring glass of wine, sorting trash, putting eggs in it's tray etc.   https://lnkd.in/gSsgUQ3C     4) Etched introducing Oasis: the first payable AI-generated game of minecraft. No coding no game engine, no logic built. I believe future of the gaming would be AI generated games instead of hard coded logic of the games. It will allow to personalize and provide unique gaming experience to each player.   (10) Etched (@Etched) / X   5) Meta released advanced embodied AI that provides touch perception, dexterity and human-robot interaction. It allows robot to feel with touch sensory allowing it to grab things with right force depending on the object type.   6) Sundar Pichai said on the earnings call that more than 25% of all new code at Google is now generated by AI. Though AI didn't generate end-to-end code for a project or product, but generated codes for functions, modules etc. I believe eventually AI will reach to intelligence level that will generate end-to-end code for product or project. 

𝐆𝐚𝐦𝐞 𝐂𝐡𝐚𝐧𝐠𝐞𝐫? 𝐆𝐨𝐨𝐠𝐥𝐞'𝐬 𝐆𝐞𝐦𝐢𝐧𝐢 1.5 𝐀𝐢𝐦𝐬 𝐭𝐨 𝐒𝐮𝐫𝐩𝐚𝐬𝐬 𝐎𝐩𝐞𝐧𝐀𝐈'𝐬 𝐂𝐡𝐚𝐭𝐆𝐏𝐓 Somewhat lost in the hype and hoopla surrounding OpenAI's launch of Sora (Text to Video Generator https://lnkd.in/gRTfATSD ) was Google DeepMind's unveiling of Gemini 1.5, the next generation of its large language model with groundbreaking capabilities in long-context understanding. The new model boasts significant performance enhancements and a revolutionary feature: the ability to process and reason about vast amounts of information within a single prompt. 𝐊𝐞𝐲 𝐢𝐦𝐩𝐫𝐨𝐯𝐞𝐦𝐞𝐧𝐭𝐬: 𝐄𝐧𝐡𝐚𝐧𝐜𝐞𝐝 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞: Gemini 1.5 outperforms its predecessor on 87% of benchmarks and matches the performance of the larger 1.0 Ultra model. 𝐇𝐢𝐠𝐡𝐥𝐲 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐭 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞: The model leverages Mixture-of-Experts (MoE) technology for faster and more affordable training and deployment. 𝐁𝐫𝐞𝐚𝐤𝐭𝐡𝐫𝐨𝐮𝐠𝐡 𝐥𝐨𝐧𝐠-𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐮𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠: Process up to 1 million tokens (equivalent to 700,000 words or 1 hour of video) at once, enabling deeper analysis and reasoning. 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬: Seamlessly analyze and understand text, code, images, audio, and video within a single prompt. 𝐄𝐱𝐭𝐞𝐧𝐬𝐢𝐯𝐞 𝐞𝐭𝐡𝐢𝐜𝐬 𝐚𝐧𝐝 𝐬𝐚𝐟𝐞𝐭𝐲 𝐭𝐞𝐬𝐭𝐢𝐧𝐠: Google DeepMind remains committed to responsible AI development with rigorous testing procedures in place. 𝐄𝐚𝐫𝐥𝐲 𝐚𝐜𝐜𝐞𝐬𝐬 𝐚𝐧𝐝 𝐟𝐮𝐭𝐮𝐫𝐞 𝐫𝐨𝐚𝐝𝐦𝐚𝐩: Developers and enterprise customers can participate in a limited preview of Gemini 1.5 Pro through AI Studio and Vertex AI. A wider release with a standard 128,000 token context window is approaching. Early testers can experiment with the 1 million token window for free during the testing period (latency times may be longer). The battle of the LLMs just heated up further! For more #ai related insights and news please subscribe to my newsletter 𝐃𝐄𝐄𝐏𝐚𝐤𝐀𝐈: 𝐀𝐈 𝐃𝐞𝐦𝐲𝐬𝐭𝐢𝐟𝐢𝐞𝐝! (https://lnkd.in/gHEjMSA5). Thanks. #DEEPakAI #AI #MachineLearning #NaturalLanguageProcessing #LargeLanguageModels #Innovation Google Google DeepMind OpenAI #chatgpt #geminiai #googledeepmind #openai https://lnkd.in/gmv6Xz2K

Google is stepping up its game in the race for advanced AI, reportedly developing a "thinking" AI to rival OpenAI’s o1 models. This move signals a significant shift as Google aims to enhance its problem-solving capabilities through chain-of-thought prompting on its Gemini platform. With OpenAI demonstrating impressive results in complex tasks like mathematics and coding, Google's reaction underscores the urgency to regain ground in the AI landscape. However, this advancement isn’t without challenges. Increased computational demands could lead to higher operational costs, and pricing strategies will be crucial as users weigh their options. As the battle for AI supremacy heats up, it's fascinating to see how companies adjust their strategies. Google’s focus on human-like reasoning may well redefine the standards of intelligence in AI. Will they catch up? Only time will tell, but the stakes have never been higher. What's your take on this competitive landscape?

Yesterday was a big day in AI! As I was attending Daniel W. Linna Jr.'s excellent AI and Law symposium at Northwestern University, people were digesting major announcements by Google and OpenAI. The previous day, OpenAI announced (but has not yet released) Sora, a model that can create realistic video from text instructions, and Google DeepMind (coincidentally?) released the Gemini 1.5 Pro model, which will have a whopping million token context window (though the initial release only supports 128K tokens) and can ingest an entire movie. Now that we have the technology, I think my next project will be to have Gemini dictate a movie to Sora, then ingest that movie and write the sequel, repeating ad nauseam in an endless loop of gripping entertainment that will last viewers a lifetime 😵 You can find overviews of Gemini 1.5 and Sora here: Gemini 1.5 - https://lnkd.in/gwsxABWK Sora - https://openai.com/sora For those who want to dig deeper, here are links to the respective technical reports: Gemini 1.5 - https://lnkd.in/gH_J-9hj Sora - https://lnkd.in/gE4FUKcD The technical report on Sora is light on actual technical details, so calibrate your expectations accordingly. #openai #google #deepmind #sora #geminipro #multimodalai

Maybe because they named it Gemini 1.5 Pro Experimental 0801, maybe because it is only accessible as a chatbot via the Google AI studio, or maybe because of all of the other drama, but LinkedIn hasn’t been noticing how good the newest Google model is, even though it tops the current AI leaderboards. Worth some experimentation, especially as it can take in long video clips, documents, etc. Only way to play with it is here, apparently: https://lnkd.in/e2vjb6b9

Google's New Gemini 1.5 Pro Experimental. I have to say I tried the previous Google Gemini's and so far have not been impressed. Currently, I'm with Anthropic 3.5 Sonnet (Best Linguistic Abilities) and impressive coding (Tableau level intereactive visualizations for business analytics and research) and then secondary Open AI GPT4o (Omni Multimodal). I'm willing to try again but frankly not expecting much anymore. If anyone does try this out comparatiely, please send me a note or comment on this post. I"m particularly looking at linguistic acuity and also more impressive multimodal abilities than the basics as between them OpenAI and Anthropic currently have this covered.

Despite all the talk of slowing AI progress, models keep leapfrogging. Google's Gemini Exp-1121 is now #1 almost across the board. A few thoughts on the latest Chatbot Arena leaderboard results: 📈 Our baseline expectations for AI advances has become the staggering pace of the last two years. If it is slower, that doesn't mean that it has stopped, or even that it is slow, just possibly not breathtaking. 🐸 The last year has seen rapid leapfrogging of models on the leaderboards, with OpenAI, Anthropic, and Google trading places at the top, and lots of jostling between the top open source models. Every leapfrog is an advance in capabilities. ✨ OpenAI's o1-preview will be followed by the full o1 model, which may be the same as its Orion model supposed to be released next month. Anthropic's Dario Amodei has not committed to its mooted lead model Claude Opus 3.5 being released, but it is likely. Llama 4 is due out soon. 🪐 Grok is already doing very well, with xAI being on the verge of releasing a consumer app (news just out - link in comments) and putting its record-breaking build of 100K GPUs into production. ⚡Chinese open source model Yi-Lightning (from Kai Fu Lee's O1.ai) is doing very well, with many other Chinese open source models such as Qwen and Deepseek performing strongly a little bit down the table. 📉 Claude Sonnet 3.5 - recently at the top of the leaderboard - is now just tenth, showing how fast things move. It's a fabulous model in many ways, but on some measures, many are beating it. Note that if you want to give the top Gemini model a whirl you can do it for free in Google's AI Studio. More digging into the pace of LLM advances coming up.

Great to watch OpenAI's latest announcement of gpt-4o! Recommend watching it (available at openai.com), but some key takeaways: 𝗛𝗮𝗹𝗳 𝗰𝗼𝘀𝘁 + 𝗱𝗼𝘂𝗯𝗹𝗲 𝘀𝗽𝗲𝗲𝗱 + 𝗮𝗰𝗰𝗲𝘀𝘀 𝘁𝗼 𝗳𝗿𝗲𝗲 𝘂𝘀𝗲𝗿𝘀 = 𝗔𝗜 𝗔𝗰𝗰𝗲𝗹𝗲𝗿𝗮𝘁𝗶𝗼𝗻 This continues the race to the bottom (in a good way) of making great AI ubiquitous, and makes it easier and cheaper to build products with AI. If it wasn't already, AI will be everywhere, and soon. Personally, this halves my product's cost, which frees me to do more to improve the value proposition. 𝗩𝗼𝗶𝗰𝗲 𝗮𝗻𝗱 𝘃𝗶𝗱𝗲𝗼 𝗶𝗻 𝗿𝗲𝗮𝗹𝘁𝗶𝗺𝗲 = 𝗕𝗶𝗴 𝗣𝗼𝘁𝗲𝗻𝘁𝗶𝗮𝗹 This opens up a ton of really cool capabilities! The demo wasn't the smoothest here, but a ton of potential. If you haven't tried voice interaction yet with an AI, it's surprisingly impactful. GPT-4o will make that even more so. 𝗗𝗲𝘀𝗸𝘁𝗼𝗽 𝗮𝗽𝗽 + 𝗿𝗲𝗮𝗹𝘁𝗶𝗺𝗲 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝗶𝗲𝘀 = 𝗬𝗼𝘂𝗿 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 They got to this at the end, but being able to screenshare with the AI and discuss things live could also be really big. You could envision a future where you have an AI passively observing your screen and trying to pre-emptively help you. Terrifying if used for surveillance by Corps/Gov't, but incredibly powerful for users.