Strawberry AI Model
Highlights
1. OpenAI has launched “o1-preview,” a new AI model designed for complex reasoning tasks, signaling a shift toward more autonomous AI agents.
2. Salesforce has launched its autonomous AI platform, Agentforce, allowing businesses to deploy AI agents.
3. Apple Intelligence is Apple’s new AI feature integrated into iPhone 16, offering advanced writing, image creation, and Siri enhancements.
4. DeepSeek-v2.5, a new open-source LLM, is reported to have superior performance in benchmarks.
5. SambaNova Systems has launched SambaNova Cloud, which is claimed to be the world’s fastest AI platform, featuring their Llama 3.1 405B model.
6. Chai-1 is a new multi-modal foundation model for molecular structure prediction that excels in various drug discovery tasks.
7. Tesla has launched its first all-electric “Giga Train” in Germany.
Innovation Insights
1. How to Manage Breakthrough Innovation (Harvard Business Review)
Astro Teller, Captain of Moonshots at X (Alphabet’s R&D division), discusses the structured approach his team uses to innovate and make decisions on ambitious projects. The team filters ideas by testing them quickly and determining whether they are worth further investment, aiming to fail fast and learn efficiently. Key traits they seek in talent include fearlessness, humility, teamwork, and a growth mindset. Teller emphasizes the importance of “monkeys and pillars,” a concept that advocates tackling the hardest parts of a project first to assess feasibility.
2. Can GenAI do your next strategy task? Not yet. (California Management Review)
The article examines whether GenAI can independently perform strategic management tasks, concluding that it is not yet fully capable. It identifies key challenges in automating strategic tasks: multi-step reasoning, context-dependence, and understanding human behavior. Experiments conducted with ChatGPT-4 on tasks like market research and strategy evaluation showed that while GenAI can automate large-scale synthesis tasks, it struggles with complex, multi-step reasoning and quantitative tasks without human guidance.
3. About OpenAI’s “Strawberry”
OpenAI’s new “Strawberry” AI model, o1-preview, demonstrates advanced reasoning capabilities by thinking through complex problems, allowing it to tackle tasks like challenging physics problems and crossword puzzles that require iterative planning and revision, although it still has some limitations and errors. This development signals a shift toward more autonomous AI agents, prompting questions about how humans will adapt their collaboration with increasingly capable AI systems (One Useful Thing).
OpenAI has acknowledged that the o1 models, exhibiting advanced reasoning and problem-solving abilities, increase the risk of misuse, particularly in the creation of biological weapons. OpenAI’s system card rated the risk related to chemical, biological, radiological, and nuclear (CBRN) weapons as “medium,” the highest risk level they’ve given. This has prompted experts, including AI scientist Yoshua Bengio, to emphasize the need for legislation to regulate the development and deployment of such advanced AI systems (Financial Times).
AI Innovations
1. OpenAI
OpenAI has launched “o1-preview,” a new AI model designed for complex reasoning tasks in science, coding, and math, as well as a cheaper, faster coding-focused variant called “o1-mini” (OpenAI).
2. Anthropic
Anthropic has introduced Workspaces in its API Console to help developers manage multiple Claude deployments by organizing resources, setting granular spend and rate limits, and streamlining access controls (Anthropic).
3. Salesforce
Salesforce has launched its autonomous AI platform, Agentforce, which it calls “the third wave of the AI revolution,” allowing businesses to deploy AI agents that can handle complex tasks and collaborate with humans (Yahoo).
Salesforce has launched Industries AI, embedding over 100 pre-built and customizable AI capabilities into 15 industry-specific clouds to address unique tasks, such as optimizing inventory management or improving student recruitment, all backed by Data Cloud and the Einstein Trust Layer (Salesforce).
4. Apple
Apple Intelligence is Apple’s new AI feature integrated into iPhone 16, offering advanced writing, image creation, and Siri enhancements, with a limited beta rollout currently available on specific devices before its full release later in the fall (CNET).
5. Google
Google’s Gemini has introduced a new feature called Gems, allowing users to create custom AI chatbots with unique personalities and purposes tailored for specific tasks, such as learning French or planning vacations (Tom’s Guide).
Google’s DataGemma is a set of open models that help address AI hallucinations by grounding large language models (LLMs) in real-world statistical data from Google’s Data Commons, a vast repository containing over 240 billion data points from trusted sources (Google).
Google’s NotebookLM app now features an AI-generated podcast option where two AI hosts discuss and summarize research notes provided by the user (The Verge).
6. Oracle
Oracle Cloud Infrastructure announced the first zettascale OCI Supercluster, powered by NVIDIA’s latest-generation GPUs, to accelerate AI workloads and data processing, offering enterprises up to 2.4 zettaflops of peak AI compute (NVIDIA).
7. Other models
Reflection 70B, an open-source AI model touted as a top performer that hallucinates less, faced scrutiny after third-party evaluators couldn’t replicate its results, leading to accusations of inaccuracy and misrepresentation (VentureBeat).
French AI startup Mistral has launched Pixtral 12B, its first multimodal AI model capable of processing both images and text (Mashable).
Adobe Firefly is now expanding into video editing with the upcoming Firefly Video Model, which offers capabilities like text-to-video generation and the ability to create b-roll, fill timeline gaps, and extend clips, all designed to be commercially safe and available in beta later this year (Adobe).
DeepSeek-v2.5 has been released and is claimed to have superior performance in benchmarks, surpassing leading models like GPT-4 Turbo, Claude 3, and Google Gemini in tasks such as coding, mathematical reasoning, and creative writing (Geeky Gadgets).
8. Biomedical use of AI
Chai-1 is a new multi-modal foundation model for molecular structure prediction that excels in various drug discovery tasks, achieving high success rates in benchmarks like PoseBusters (Chai Discovery).
The Virchow foundation model, developed by Paige and Microsoft Research, is an advanced AI tool for cancer detection that excels in identifying rare and complex cancers, offering 94% accuracy in diagnosing previously hard-to-detect cancers (Healthcare IT News).
9. Others
A company called Altera.ai created an experiment called Project Sid, in which 1,000 autonomous AI agents were given access to a Minecraft world, where they quickly formed alliances, created a currency system, and developed complex social behaviors (Tom’s Guide).
Uber is partnering with Waymo to offer driverless rides in Austin and Atlanta starting in 2025, using Waymo’s fully autonomous, all-electric Jaguar I-Pace vehicles, with Uber handling tasks like vehicle cleaning and repairs (CNET).
SambaNova Systems has launched SambaNova Cloud, the world’s fastest AI platform, featuring their Llama 3.1 405B model achieving a record speed of 132 output tokens per second, outperforming models from OpenAI, Anthropic, and Google (Business Wire).
Hugging Face has introduced LightEval, an open-source evaluation suite designed to help companies and researchers assess LLM with a focus on transparency, customization, and ethical standards (VentureBeat).
YouTube Music’s new AI-powered feature, “Ask Music,” allows users to create custom radio stations based on specific musical preferences, such as genres or characteristics like “melancholy indie rock” or “jangly 60s pop” (TechRadar).
Hume’s EVI 2 is a speech-to-speech AI voice assistant that builds on its previous version with more natural-sounding voice and emotional understanding, offering features like sub-second response times, tone adaptation, and speech modulation (Tom’s Guide).
The newly announced “FiveThirtyNine” AI forecasting bot, built on GPT-4o, provides superhuman-level predictions for various queries by searching for relevant news and opinion articles, compiling facts, and adjusting for biases to deliver calibrated probabilities (Safe AI).
Infineon Technologies announced a breakthrough in producing 300mm gallium nitride (GaN) wafers, which can produce 2.3 times more chips than the current 200mm wafers, potentially lowering costs and accelerating AI applications (Yahoo).
Other Innovations
1. Quantum computing
Google claims a quantum computing breakthrough with a surface code technique that reduces errors in quantum bits, potentially paving the way for more reliable and practical quantum computers (MIT Technology Review).
2. Building design
Neuroscientists and architects are using a large laboratory in East London to create life-size simulations of real-world environments, studying how people navigate and respond to these spaces, with the aim of improving future building designs for a variety of needs (MIT Technology Review).
3. Transparent skin
Researchers used the food dye tartrazine (Yellow No. 5), also used in Doritos, to temporarily make the skin of living mice transparent, allowing them to view internal organs and structures without invasive procedures (Scientific American).
4. Electric train
Tesla has launched its first all-electric “Giga Train” in Germany, offering free rides and capable of transporting up to 500 passengers with plans to eventually accommodate 4,500 employees (Yahoo).
5. Trifold phone
Huawei has launched the Mate XT Ultimate Design, the world’s first dual-hinged, triple-screen foldable phone (The Verge).
6. Internet Archive
Google Search is now adding links to the Internet Archive’s Wayback Machine in its results, allowing users to access archived versions of webpages, which serves as an alternative to Google’s recently removed cached pages feature (The Verge).
7. Robots
Google’s robotics team has introduced two AI systems, ALOHA Unleashed and DemoStart, which significantly enhance robot dexterity, enabling them to perform complex tasks like tying shoelaces and tightening bolts using two arms or multi-fingered hands (Google).
