TechPrufer
Get started
Google
Google logoGoogle · Feature trail

Predictive APIs and Machine Learning Integration

64 posts · Sep 2010 – Jul 2026

What this trail covers

The Gemini API continues to evolve with updated production-ready models (Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002), featuring improved quality in math, long context, and vision benchmarks. Pricing for 1.5 Pro has been significantly reduced, and rate limits for both 1.5 Flash and Pro have been increased. Latency has been reduced, and output speed has been increased. Default filter settings are now configurable by developers. An experimental update to Gemini-1.5-Flash-8B is also available. The `logprobs` feature has been officially introduced in the Gemini API on Vertex AI, allowing developers to see the probability scores for the chosen token and its top alternatives, providing insight into the model's reasoning. This feature can be applied to use cases like confident classification, dynamic autocomplete, and quantitative RAG evaluation.

All Google trails
  1. Building scalable AI agents with modular prompt transpilation- Google Developers BlogThis post introduces the concept of 'modular prompt transpilation' as a solution to the scaling problem of monolithic AI agent prompts. It advocates for treating prompts as build artifacts, using modular skill files, and implementing build-time validation and CI/CD integration to ensure prompt reliability and maintainability. It also explores dynamic skill loading and the potential for agents to propose their own improvements through a structured code change process.
  2. Evolving Spec-Driven Development: Conductor Now Supports Antigravity- Google Developers BlogThis post announces the evolution of Conductor from a Gemini CLI extension to a portable plugin, enabling spec-driven development through conversational interfaces. It highlights improved compatibility with tools like Antigravity CLI, natural interaction, and preserved context, making AI-assisted development more accessible and adaptable.
  3. Expanding Choice in Gemini Enterprise Agent Platform: Introducing Grounding with Parallel Web Search- Google Developers BlogThis post introduces Parallel Web Systems as a new, natively integrated web grounding provider for the Gemini Enterprise Agent Platform. This expands developer choice and flexibility by allowing Gemini models to be grounded in high-quality, real-time web search results with exact citations, enhancing the accuracy and verifiability of AI agents. It enables new use cases such as catalog enrichment, autonomous agents for compliance checks, and multi-agent orchestration by allowing programmatic calls, data caching, and post-processing of search results.
  4. LiteRT.js, Google's high performance Web AI Inference- Google Developers BlogIntroduces LiteRT.js, a JavaScript binding for LiteRT, enabling high-performance AI inference directly within web browsers via WebAssembly. It leverages native hardware acceleration (XNNPACK, ML Drift, WebNN) for CPU, GPU, and NPU, offering up to 3x speedup over existing web runtimes. It also details PyTorch conversion with LiteRT Torch and tailored quantization with AI Edge Quantizer, and showcases real-world applications like YOLO integration, depth estimation, and image upscaling.
  5. Gemma 4 12B: The Developer Guide- Google Developers BlogIntroduced Gemma 4 12B, a dense multimodal model with a unified, encoder-free architecture. This model features reduced multimodal latency by bypassing separate vision and audio encoders, is the first medium-sized Gemma model with native audio input, and is designed for local inference on GPUs with 16GB VRAM. A new macOS desktop experience is also released. The architecture utilizes a single decoder-only transformer with a vision embedder (35M parameters) and linear projection for audio wave input. Unified fine-tuning allows for updating the entire multimodal token loop. The model demonstrates capabilities in automatic speech recognition, agentic reasoning, diarization, video understanding, and coding. On-device and desktop serving are powered by LiteRT-LM, including native macOS apps and OpenAI-compatible local API servers. Getting started resources include LM Studio, Ollama, Google AI Edge Gallery App, Google AI Edge Eloquent app, LiteRT-LM CLI, Hugging Face, Kaggle, developer documentation, quick start notebooks, and integrations with Transformers, llama.cpp, MLX, SGLang, vLLM, and Unsloth. A Gemma Skills Repository is released to support agentic development.
  7. Supercharging LLM inference on Google TPUs: Achieving 3X speedups with diffusion-style speculative decoding- Google Developers BlogThis post details the successful implementation of DFlash, a diffusion-style speculative decoding method, on Google TPUs within the vLLM inference framework. Key technical contributions include the 'dual-cache' solution for attention compatibility with paged attention, intelligent context management using power-of-2 padding for efficient host-TPU communication, and bridging the metadata gap in TPU inference to ensure strict alignment between the draft and target models. Benchmarking demonstrated an average 3.13x speedup on TPU v5p, with peak gains of nearly 6x on math tasks, and a 2.29x end-to-end serving speedup compared to EAGLE-3. The 'K-Flat' breakthrough, showing minimal cost for verifying larger token blocks on TPUs, is also highlighted.
  8. Build a smart financial assistant with LlamaParse and Gemini 3.1- Google Developers BlogThis post introduces the integration of LlamaParse with Gemini 3.1 Pro and Gemini Flash to build an intelligent personal finance assistant. It details a four-stage workflow: ingest, route, extract (text and tables in parallel), and synthesize. The post highlights the use of Gemini 3.1 Pro for layout comprehension and Gemini Flash for summarization, along with the setup and execution of the workflow using LlamaIndex Workflows.
  9. Data Commons Hosted MCP: Zero-Install Public Data for AI- Google Developers BlogIntroduced a hosted Data Commons MCP service on Google Cloud Platform, removing the need for local Python environments and infrastructure management for AI agents interacting with Data Commons data. This enhances scalability and simplifies setup for users of the Gemini CLI extension and other agents.
  10. Announcing the Data Commons Gemini CLI extension- Google Developers BlogIntroduces a new Gemini CLI extension for Data Commons, enabling natural language querying of public datasets. This extension leverages Data Commons MCP tools for data exploration, analysis, and grounding LLM responses in authoritative sources. It allows for comparison with other Gemini CLI tools and integration into broader workflows with other data extensions.
  11. New Gemini API updates for Gemini 3- Google Developers BlogIntroduced Gemini 3 via the Gemini API with updates including: a new `thinking_level` parameter for controlling model reasoning depth, a `media_resolution` parameter for granular control over multimodal vision processing, "Thought Signatures" to maintain reasoning chains for agentic workflows and image generation, and the ability to combine Grounding with Google Search and URL context for structured outputs. Updated pricing for Grounding with Google Search to a usage-based model. Provided best practices for using Gemini 3 Pro and a System Instructions template for agentic coding.
  12. Unlocking Multi-Spectral Data with Gemini- Google Developers BlogIntroduces a technique for analyzing multi-spectral imagery using Gemini's multimodal capabilities by mapping invisible spectral bands (NIR, SWIR) to RGB channels to create 'false-color composite' images. This allows developers to leverage Gemini's reasoning engine for remote sensing tasks without specialized models, demonstrated with examples from the EuroSat dataset and a Colab notebook for practical application.
  13. Gemma explained: EmbeddingGemma Architecture and Recipe- Google Developers BlogThis post details the architecture and training recipe of EmbeddingGemma, a new text embedding model derived from Gemma 3. It explains how text is converted into embeddings using a transformer, pooling, and dense layers, and describes the three weighted loss functions (NCE, GOR, Geometric Embedding Distillation) used for training. It also introduces Matryoshka Representation Learning (MRL) for flexible embedding dimensionality and outlines the development recipe including encoder-decoder training, pre fine-tuning, fine-tuning, model souping, and Quantization-Aware Training (QAT).
  14. Continuing to bring you our latest models, with an improved Gemini 2.5 Flash and Flash-Lite release- Google Developers BlogThis post announces the release of updated versions of Gemini 2.5 Flash and 2.5 Flash-Lite. Gemini 2.5 Flash-Lite now offers better instruction following, reduced verbosity (leading to cost and latency improvements), and stronger multimodal and translation capabilities. Gemini 2.5 Flash has improved agentic tool use (5% gain on SWE-Bench Verified) and is more cost-efficient. New `-latest` aliases (`gemini-flash-latest`, `gemini-flash-lite-latest`) are introduced for easier access to the most recent model versions.
  15. Gemini Batch API now supports Embeddings and OpenAI Compatibility- Google Developers BlogIntroduced support for the Gemini Embedding model within the Gemini Batch API, enabling asynchronous embedding generation at higher rate limits and lower costs. Added OpenAI SDK compatibility for the Gemini Batch API, allowing developers to use the OpenAI SDK to submit and process batches for chat completions.
  16. Veo 3 and Veo 3 Fast – new pricing, new configurations and better resolution- Google Developers BlogIntroduced Veo 3 and Veo 3 Fast models as stable, production-ready capabilities within the Gemini API. Added support for vertical video output (9:16 aspect ratio) and 1080p HD resolution. Significantly reduced pricing for both Veo 3 and Veo 3 Fast. Provided a Python code example demonstrating video generation with the new parameters and highlighted developer use cases.
  17. From Fine-Tuning to Production: A Scalable Embedding Pipeline with Dataflow- Google Developers BlogThis post introduces a scalable, real-time knowledge ingestion pipeline for AI embeddings using Google Cloud Dataflow and the open-source EmbeddingGemma model. It details how to leverage Dataflow's MLTransform to integrate EmbeddingGemma for efficient, self-contained embedding generation and subsequent writing to vector databases like AlloyDB. The post highlights the benefits of data locality, unified system management, and scalability offered by this approach, and provides code examples for configuring the embedding transform and writing to AlloyDB.
  18. Introducing EmbeddingGemma: The Best-in-Class Open Model for On-Device Embeddings- Google Developers BlogIntroduces EmbeddingGemma, a new open embedding model based on Gemma 3 architecture, optimized for on-device AI. It delivers best-in-class performance for its size (308M parameters), supporting over 100 languages and enabling RAG and semantic search on-device. Key features include flexible output dimensions via Matryoshka Representation Learning, low RAM usage (<200MB) with quantization, and fast inference times (<15ms on EdgeTPU). It integrates with popular tools like sentence-transformers, llama.cpp, and LangChain.
  19. Stop “vibe testing” your LLMs. It's time for real evals.- Google Developers BlogIntroduces Stax, an experimental developer tool designed to streamline the LLM evaluation lifecycle. Stax allows developers to build custom benchmarks and autoraters (LLM-as-a-judge) to rigorously test their AI stacks, moving beyond subjective 'vibe testing' to data-driven decision-making. It supports bringing or building datasets, offers out-of-the-box autoraters for common checks, and enables custom autorater creation based on specific criteria like brand voice or application rules.
  20. URL context tool for Gemini API now generally available- Google Developers BlogThe URL context tool for the Gemini API is now generally available for scaled production use. It has expanded content support to include PDFs, images (PNG, JPEG, BMP, WebP), and standard web pages (HTML), structured data (JSON, XML, CSV), and text files (Plain Text, RTF, CSS, JavaScript). Rate limits are now model-specific, and pricing is based on added input tokens. This enables more powerful generative AI applications by allowing models to ingest and reason over the full content of specific webpages, going beyond search snippets.
  21. Introducing Gemma 3 270M: The compact model for hyper-efficient AI- Google Developers BlogIntroduced Gemma 3 270M, a compact 270-million parameter model designed for task-specific fine-tuning. It features a large vocabulary (256k tokens) for handling specific tokens, extreme energy efficiency (0.75% battery usage on Pixel 9 Pro SoC for 25 conversations with INT4 quantization), strong instruction-following capabilities out-of-the-box, and production-ready Quantization-Aware Trained (QAT) checkpoints for INT4 precision deployment on resource-constrained devices. The model is positioned as the 'right tool for the job' for high-volume, well-defined tasks, cost-sensitive applications, rapid iteration, user privacy on-device, and building fleets of specialized models.
  22. Veo 3 Fast and new image-to-video capabilities- Google Developers BlogIntroduces Veo 3 Fast, an optimized model for speed and price, offering text-to-video and image-to-video capabilities. Veo 3 Fast is suitable for programmatic advertising, rapid prototyping, and content creation at scale, priced at $0.40/second with audio. New image-to-video capabilities are also available for both Veo 3 and Veo 3 Fast, allowing users to transform still images into video clips with sound, maintaining consistency from the input image. These capabilities are accessible via the Gemini API and are priced the same as text-to-video outputs ($0.75/second with audio for Veo 3).
  23. Gemini Embedding: Powering RAG and context engineering- Google Developers BlogThis post details the adoption and performance of the Gemini Embedding model (`gemini-embedding-001`) by various organizations (Box, re:cap, Everlaw, Roo Code, Mindlid, Interaction Co.) for use cases including retrieval-augmented generation (RAG), context engineering, document analysis, financial transaction classification, legal discovery, codebase search, and AI assistant context. It highlights performance metrics such as increased recall, F1 score improvements, accuracy rates, reduced latency, and faster embedding times compared to previous models and competitors. The Matryoshka property of Gemini Embedding is also mentioned for its efficiency benefits.
  24. Introducing LangExtract: A Gemini powered information extraction library- Google Developers BlogIntroduces LangExtract, an open-source Python library that leverages Gemini models for information extraction from unstructured text. Key features include precise source grounding, reliable structured outputs via controlled generation, optimized long-context extraction, interactive visualization, flexible LLM backend support, and domain flexibility through few-shot learning. The library also supports inferring supplementary information using LLM world knowledge.
  25. Gemini 2.5 Flash-Lite is now stable and generally available- Google Developers BlogThis post announces the stable and general availability of Gemini 2.5 Flash-Lite, a new model in the Gemini 2.5 family. It highlights its cost-efficiency ($0.10 input per 1M, $0.40 output per 1M), speed (lower latency than 2.0 Flash-Lite and 2.0 Flash), quality (higher than 2.0 Flash-Lite across benchmarks), and features (1 million-token context window, controllable thinking budgets, Grounding with Google Search, Code Execution, and URL Context). It also provides examples of its use in production by Satlyt (45% latency reduction, 30% power consumption decrease), HeyGen (video translation into over 180 languages), DocsHound (processing long videos and extracting thousands of screenshots), and Evertune (fast analysis and report generation for model output insights).
  26. Conversational image segmentation with Gemini 2.5- Google Developers BlogIntroduces conversational image segmentation capabilities for Gemini 2.5, allowing users to query images using natural language. This feature supports object relationships, conditional logic, abstract concepts, in-image text recognition (OCR), and multi-lingual labels. It enables new use cases in interactive media editing, safety monitoring, and insurance claims assessment. Developers can leverage this through a single API, reducing the complexity of building vision applications.
  27. Build with Veo 3, now available in the Gemini API- Google Developers BlogIntroduces Veo 3, a new video generation model available via the Gemini API and Vertex AI. Highlights its capabilities in generating high-fidelity video with native audio, realistic physics, and cinematic quality. Provides examples of its use by Cartwheel and Volley. Includes a Python code snippet for generating videos using the Gemini API and mentions SynthID watermarking for responsible AI development.
  28. Unlock Gemini’s reasoning: A step-by-step guide to logprobs on Vertex AI- Google Developers BlogIntroduces the `logprobs` feature for the Gemini API on Vertex AI, enabling developers to access token probability scores and top alternative token probabilities. Provides code examples for enabling and processing logprobs, and demonstrates use cases in classification (ambiguity detection, confidence-based thresholding), dynamic autocomplete, and quantitative RAG evaluation. Explains the mathematical concept of log probabilities and their relation to model confidence.
  29. Gemini Embedding now generally available in the Gemini API- Google Developers BlogAnnounces the general availability of the `gemini-embedding-001` text model in the Gemini API and Vertex AI. Highlights its top performance on the MTEB Multilingual leaderboard, its versatility across domains, and its technical details including multi-language support (100+ languages), 2048 token input length, and Matryoshka Representation Learning (MRL) for scalable output dimensions. Details pricing ($0.15 per 1M input tokens) and rate limits, with free and paid tiers available. Provides a Python code example for using the `embed_content` endpoint and links to documentation and quickstart notebooks. Notifies developers about the deprecation of experimental and legacy embedding models.
  30. Batch Mode in the Gemini API: Process more for less- Google Developers BlogIntroduces a new asynchronous batch mode for the Gemini API, designed for high-throughput, non-latency-critical workloads. This mode offers a 50% cost reduction and higher rate limits compared to synchronous APIs. The workflow involves packaging requests into a single file, submitting it, and retrieving results within 24 hours. Examples include bulk content generation/processing (e.g., video analysis by Reforged Labs) and model evaluations (e.g., by Vals AI). A Python SDK example demonstrates its usage.
  31. Imagen 4 is now available in the Gemini API and Google AI Studio- Google Developers BlogIntroduces Imagen 4, Google's latest text-to-image model, available via the Gemini API and Google AI Studio. Highlights improved text rendering and generation quality compared to prior models. Details two model variants: Imagen 4 (general use) and Imagen 4 Ultra (precision). Provides pricing information and links to documentation and cookbooks. Mentions SynthID watermarking for trust.
  32. Exploring the Magic Mirror: an interactive experience powered by the Gemini models- Google Developers BlogThis post details the 'Magic Mirror' project, an interactive experience powered by the Gemini models and the JavaScript GenAI SDK. It highlights the use of the Gemini API's Live API for real-time, bidirectional audio streaming and conversational interactions, including interruption handling. It also showcases the application of Function Calling for integrating external tools like image generation, grounding with Google Search for factual information retrieval, system instructions for shaping AI tone, speech configuration for voice customization, and modality control for output flexibility (text, audio, etc.). The post provides links to the project's code on GitHub and a tutorial on Hackster.io.
  33. Gemini API I/O updates- Google Developers BlogThis post details several updates to the Gemini API and Google AI Studio. New models include Gemini 2.5 Flash preview (gemini-2.5-flash-preview-05-20) with improved reasoning, code, and long context capabilities, alongside cost-efficiency gains. Gemini 2.5 Pro and Flash text-to-speech (TTS) previews offer native audio output, speaker control, and multispeaker support across 24 languages. Gemini 2.5 Flash native audio dialog (in preview) provides natural sounding voices, proactive audio distinction, and emotional tone response. Lyria RealTime is now available for live music generation via WebSockets. Gemini 2.5 Pro Deep Think is an experimental reasoning mode for complex math and coding. Gemma 3n is introduced as an open model optimized for edge devices, supporting text, audio, and vision inputs with parameter-efficient processing. API functionality enhancements include thought summaries for 2.5 Pro and Flash for debugging, thinking budgets for 2.5 Flash to control model thinking, a new experimental URL context tool for retrieving information from links, and a computer use tool for browser control capabilities. Structured output support is improved with broader JSON Schema support, including '$ref' and tuple-like structures. Video understanding improvements allow YouTube video URLs or uploads for summarization, translation, and analysis, with support for video clipping, dynamic FPS, and multiple resolutions. Asynchronous function calling with non-blocking behavior is now supported in the Live API. A new Batch API is being tested for cost-effective, high-throughput request processing.
  34. Build and train a recommender system in 10 minutes using Keras and JAX- Google Developers BlogIntroduced Keras Recommenders (KerasRS), a new library for building state-of-the-art recommendation systems. KerasRS provides APIs with building blocks for ranking and retrieval tasks, supporting JAX, TensorFlow, and PyTorch backends. It includes specialized layers, losses, and metrics for recommender tasks, demonstrated with a `SequentialRetrievalModel` example using `BruteForceRetrieval` and `PairwiseHingeLoss`. Future plans include `DistributedEmbedding` for TPUs and continuous addition of model implementations.
  35. Start building with Gemini 2.0 Flash and Flash-Lite- Google Developers BlogIntroduces Gemini 2.0 Flash and Flash-Lite models, highlighting their improved performance, cost-effectiveness, and extended context windows. Provides examples of their application in voice AI (Daily.co), data analytics (Dawn), and video editing (Mosaic), showcasing specific use cases and benefits such as reduced search times, lower costs, and faster video editing.
  36. Vertex AI RAG Engine: A developers tool- Google Developers BlogIntroduces Vertex AI RAG Engine, a managed orchestration service for building grounded generative AI applications. It simplifies data retrieval and LLM integration, offering ease of use, customization options (parsing, chunking, annotation, embedding, vector storage, open-source models), high-quality Google components, and flexible integration with vector databases like Pinecone, Weaviate, or Vertex AI Vector Search. It positions RAG Engine as a balance between Vertex AI Search (fully managed) and fully DIY RAG approaches. Provides use cases in financial services, healthcare, and legal, along with getting started resources (notebooks, documentation, integrations, evaluation framework).
  37. Enhance your prompts with Vertex AI Prompt Optimizer- Google Developers BlogIntroduced Vertex AI Prompt Optimizer, a new service in Public Preview that automates prompt engineering for LLMs. It uses an iterative LLM-based optimization algorithm with an optimizer and evaluator model to generate and evaluate candidate prompts, selecting the best instructions and demonstrations based on user-defined metrics. This service is based on Google Research's paper on automatic prompt optimization (APO) methods and aims to save time and effort in prompt engineering, facilitate prompt migration between models, and enhance the quality of GenAI applications. The post details how to use the Vertex AI SDK for Python to prepare prompt templates, upload labeled samples, configure optimization settings, and run optimization jobs.
  38. Updated production-ready Gemini models, reduced 1.5 Pro pricing, increased rate limits, and more- Google Developers BlogThis post announces the release of updated production-ready Gemini models (Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002). Key improvements include over 50% price reduction on 1.5 Pro tokens, doubled rate limits for 1.5 Flash and tripled for 1.5 Pro, 2x faster output, and 3x lower latency. Benchmark improvements are noted in MMLU-Pro (~7%), MATH and HiddenMath (~20%), and vision/code use cases (~2-7%). Default filter settings are now configurable by developers. An experimental update to Gemini-1.5-Flash-8B is also highlighted.
  39. Mastering Controlled Generation with Gemini 1.5: Schema Adherence for Developers- Google Developers BlogThis post introduces Controlled Generation for Gemini 1.5 Flash and adds 'enum' support, enabling developers to reliably generate responses that adhere to a defined schema. It also notes that controlled generation is automatically enabled when using function calling on Gemini 1.5. The post explains the importance of controlled generation for producing machine-readable data, seamless API integration, and predictable outputs. It details the 'response schema' concept, built on OpenAPI 3.0, and highlights Google's 'secret sauce' for controlled generation: minimal latency, enum support, and no data storage. Examples are provided for generating JSON recipes and classifying product condition with enum schemas.
  40. Gemini 1.5 Flash price drop with tuning rollout complete, and more- Google Developers BlogThis post announces significant price reductions for Gemini 1.5 Flash, making it more affordable for high-volume, low-latency use cases. It also announces the general availability of Gemini 1.5 Flash text tuning for all developers via the Gemini API and Google AI Studio, allowing for model customization to improve performance, reduce prompt context size, latency, and cost, while increasing accuracy. The Gemini API's language support has been expanded to over 100 languages for both 1.5 Pro and Flash models, aiming to eliminate 'language' block finish reasons. Google AI Studio access is now frictionless for Google Workspace customers. The post also highlights improvements to developer documentation, including a revamped API reference, and introduces PDF understanding capabilities for both text and vision through the Gemini API and AI Studio. Numerous improvements to Google AI Studio itself are also detailed, including UI enhancements and performance optimizations.
  41. Get started with Gemma on Ray on Vertex AI- Google Developers BlogThis post details how to use Gemma, an open family of models, with Ray on Vertex AI for supervised fine-tuning and serving. It provides a step-by-step guide on setting up a Google Cloud project, enabling necessary APIs, creating a Cloud Storage bucket and Artifact Registry repository, and configuring a Vertex AI TensorBoard instance. It outlines the process of building a custom Docker image for the Ray cluster, creating the Ray cluster itself using the Vertex AI Python SDK, and fine-tuning Gemma using Ray Train with HuggingFace Transformers. The post includes code examples for submitting training jobs via the Ray Jobs API and discusses monitoring training artifacts and performance using Cloud Storage and Vertex AI TensorBoard.
  42. Gemini 1.5 Pro 2M context window, code execution capabilities, and Gemma 2 are available today- Google Developers BlogThis post announces the general availability of the 2 million token context window for Gemini 1.5 Pro, the launch of context caching for Gemini 1.5 Pro and 1.5 Flash, and the enablement of code execution capabilities for both models via the Gemini API and Google AI Studio. It also highlights the availability of Gemma 2 in Google AI Studio and provides examples of Gemini 1.5 Flash being used in production. Text tuning for Gemini 1.5 Flash is also rolling out.
  43. How It’s Made: AI Roadtrip, a Pixel Campaign Powered by Generative AI and Fans- Google Developers BlogThis post details the technical implementation of an AI-powered campaign called 'AI Roadtrip'. It describes the use of Gemini 1.5 Pro for script generation, Imagen 2 for background image generation, and Cloud Text-to-Speech for audio generation. It also explains how these assets are composited in Unreal Engine and how human creatives interact with the system for editing and refinement. The prompt engineering techniques used for Gemini and the prompt generation process for Imagen are also described.
  44. Gemini 1.5 Pro and 1.5 Flash GA, 1.5 Flash tuning support, higher rate limits, and more API updates- Google Developers BlogThis post announces the general availability (GA) of Gemini 1.5 Pro and 1.5 Flash models, along with significant updates to their API. Key contributions include increased rate limits for Gemini 1.5 Flash (1000 RPM, no daily limit), the upcoming rollout of tuning support for Gemini 1.5 Flash (currently free), and the introduction of JSON schema mode for more structured model output. Billing is now enabled in Google AI Studio for higher rate limits. Google AI Studio also received UI updates like light mode and mobile support.
  45. Tune Gemini Pro in Google AI Studio or with the Gemini API- Google Developers BlogIntroduces Gemini Tuning for Gemini 1.0 Pro models, accessible via Google AI Studio and the Gemini API. This feature utilizes Parameter Efficient Tuning (PET) to customize models for specific tasks, improving output quality, reducing prompt context size, and lowering response latency compared to few-shot prompting. It requires fewer data points than traditional fine-tuning and is suitable for use cases like classification, information extraction, structured output generation, and critique models. The post details the user experience in Google AI Studio for creating, selecting data for, and running tuned models.
  46. Gemini 1.5: Our next-generation model, now available for Private Preview in Google AI Studio- Google Developers BlogIntroduced Gemini 1.5 Pro, a next-generation multimodal model with a Mixture-of-Experts (MoE) architecture for improved efficiency. Enabled an experimental 1 million token context window, significantly expanding the model's capacity for processing large inputs like PDFs, code repositories, and up to 1 hour of video. Added the ability to upload multiple files and query entire code repositories within Google AI Studio. Introduced easy tuning capabilities for customizing Gemini models. Integrated Gemini API with new Firebase Extensions and Project IDX. Reduced pricing for Gemini 1.0 Pro.
  47. How it's Made - Exploring AI x Learning through ShiffBot, an AI experiment powered by the Gemini API- Google Developers BlogThis post details the technical implementation of ShiffBot, an experimental AI tool for creative coding education powered by the Gemini API. It describes the use of prompt engineering, including static persona definitions and few-shot learning examples, to imbue the AI with Daniel Shiffman's teaching style. It also explains the application of semantic retrieval using text embeddings to ground the AI's responses in a corpus of educational videos and content, enabling it to surface relevant resources and provide context-aware assistance. The integration with the p5.js editor allows ShiffBot to access conversation and code context for personalized responses.
  48. YouTube Ads Creative Analysis- Google Developers BlogThis post details a machine learning-based approach for analyzing YouTube ad creatives to understand which components affect view-through rate (VTR). It outlines a 5-step process: defining business questions, extracting raw components using Google Cloud Video Intelligence API, engineering features (manually and with LLMs), modeling VTR with engineered features, and interpreting the results. The post highlights the use of LLMs for AI-based feature engineering and discusses challenges like feature interactions and data representativeness. It also suggests activation strategies using Ads Creative Studio and Video Experiments.
  49. It's time for developers and enterprises to build with Gemini Pro- Google Developers BlogThis post announces the availability of Gemini Pro and Gemini Pro Vision for developers and enterprises. Gemini Pro is accessible via the Gemini API and Google AI Studio, offering features like a 32K context window, function calling, embeddings, semantic retrieval, custom knowledge grounding, and chat functionality. It supports 38 languages and is available in 180+ countries. Gemini Pro Vision accepts text and imagery as input. SDKs are provided for Python, Android (Kotlin), Node.js, Swift, and JavaScript. Developers can transition to Vertex AI on Google Cloud for advanced customization, data control, and enterprise-grade features. Pricing details are provided, with free access currently available through Google AI Studio and Vertex AI until early next year. Gemini Ultra is slated for release early next year.
  50. How it’s Made: Interacting with Gemini through multimodal prompting- Google Developers BlogThis post demonstrates the multimodal prompting capabilities of the Gemini API, showcasing its ability to understand and reason about combinations of images and text. It provides examples of Gemini's performance in tasks such as image description, pattern recognition, spatial reasoning, sequence understanding, and tool use (generating search queries). The post also highlights Gemini's application in prototyping multimodal games and generating code snippets, illustrating its versatility in various developer workflows.
  51. Make with MakerSuite – Part 1: An Introduction- Google Developers BlogIntroduces MakerSuite, a new tool for building generative AI applications using Google's latest models. MakerSuite offers an efficient UI for prompting, translates prompts into production-ready code, and supports three prompt types: text, data, and chat. The post also announces the removal of the waitlist for MakerSuite and highlights its use by companies like LlamaIndex and Chroma.
  52. MakerSuite expands to 179 countries and territories, and adds helpful features for AI makers- Google Developers BlogMakerSuite has expanded its availability to 179 countries and territories, including Google Workspace accounts. New features include automatic text prompt optimization, a dark mode option, and the ability to import/export data using Google Sheets and CSV files for improved collaboration and time savings. The integration with the PaLM API allows for easy transition from prompt experimentation in the browser to code integration.
  53. How it’s Made: TextFX is a suite of AI tools made in collaboration with Lupe Fiasco- Google Developers BlogThis post details the development of TextFX, a suite of AI tools for creative text generation, in collaboration with Lupe Fiasco. It describes the process of observing his creative techniques, identifying opportunities for AI augmentation, and leveraging large language models (LLMs) like Google's PaLM 2 via the PaLM API. The post highlights the use of few-shot prompting and prompt engineering to codify creative tasks such as generating same-sounding phrases, similes, and acronyms. It also mentions the use of MakerSuite for rapid prototyping and the open-sourcing of the TextFX code. The core technical contribution is the application of few-shot learning and prompt engineering with LLMs to a creative domain.
  54. Using Generative AI for Travel Inspiration and Discovery- Google Developers BlogThis post introduces an open-source Generative AI template for the travel industry, demonstrating the combination of the PaLM API with Google Maps Places API for personalized travel itinerary planning. It details the implementation using few-shot prompting, a parser for extracting locations, and conversational memory management via the 'messages' field in the PaLM API. It also highlights the potential for third-party integration with booking APIs and the use of embedding services for partner data. The post emphasizes the responsibility of building AI experiences and the potential for generative AI in various industries beyond travel.
  55. Generative AI ‘Food Coach’ that pairs food with your mood- Google Developers BlogThis post introduces Generative AI Templates, specifically the MoodFood 'Food Coach' application, which demonstrates the use of the PaLM API for creative recipe generation, data formatting into JSON, and safeguarding. It details the implementation of multiple PaLM API calls with varying temperature settings to balance creativity and structure, and explains the concept of prompting with context, messages, temperature, and candidate_count. It also highlights the use of embedding services for integrating with partner data and enabling third-party integrations for features like ingredient purchasing.
  56. PaLM API & MakerSuite moving into public preview- Google Developers BlogThis post announces the public preview of the PaLM API and MakerSuite, making PaLM 2 available to developers. It highlights PaLM 2's improved reasoning and code generation, instruction-tuning for ease of use, efficiency with advanced model architecture, high token generation speeds (75+ tokens/sec), and 8k context windows. It also details integrations into the developer ecosystem via partners, Firebase, Colab, and introduces MakerSuite with Project Tailwind as an example of an AI-first notebook built with the PaLM API.
  57. PaLM API & MakerSuite: an approachable way to start prototyping and building generative AI applications- Google Developers BlogIntroduces the PaLM API and MakerSuite as an approachable way for developers to prototype and build generative AI applications. Highlights the ease of use for getting started with generative AI, abstracting away complex infrastructure.
  58. Campaign Anomaly Detector- Google Developers BlogThis post introduces the Campaign Anomaly Detector, a system designed to identify anomalous patterns in campaign performance metrics. It leverages machine learning models to analyze historical data and detect deviations from expected behavior, enabling proactive intervention to optimize campaign effectiveness. The system incorporates anomaly detection algorithms and provides actionable insights for campaign managers.
  59. Prediction Framework, a time saver for Data Science prediction projects- Google Developers BlogThis post introduces the Prediction Framework, a new tool designed to accelerate data science prediction projects by providing a standardized and efficient workflow. It aims to reduce the time spent on common tasks such as data preparation, model training, and deployment, allowing data scientists to focus more on model development and analysis.
  60. Sudoku, Linear Optimization, and the Ten Cent Diet- Google Developers BlogThis post announces the graduation of the Google Prediction API from labs, highlighting its maturity and stability. It also showcases the application of linear optimization techniques to solve complex problems like Sudoku and the 'Ten Cent Diet', demonstrating the practical utility and breadth of Google's predictive and optimization capabilities beyond simple predictions.
  61. Google Prediction API 1.5 adds enumeration, analysis, and more- Google Developers BlogVersion 1.5 of the Google Prediction API introduces new features for enumeration, analysis, and enhanced predictive modeling. This update allows for more sophisticated data analysis and the creation of more accurate predictive models.
  62. Google Prediction API: faster, easier to use, and more accurate- Google Developers BlogThis post announces the Google Prediction API, a new service that makes it easier for developers to build applications that leverage machine learning. The API offers faster, more accurate, and easier-to-use predictive models, allowing developers to integrate machine learning capabilities into their applications without requiring deep expertise in the field. It highlights the ease of use and improved accuracy compared to previous offerings.
  63. Google Prediction API graduates from labs, adds new features- Google Developers BlogThe Google Prediction API has graduated from labs and is now generally available. It has been enhanced with new features, including the ability to train models on user-provided data and to perform real-time predictions. The API also supports a wider range of machine learning algorithms and provides improved performance and scalability.
  64. Prediction API: Tunable predictive models- Google Developers BlogThis post introduces the Prediction API, a new service that allows developers to build and deploy tunable predictive models. It provides a flexible framework for creating custom machine learning models and integrating them into applications. The API supports various model types and offers features for training, evaluation, and deployment.
  65. Prediction API: Every app a smart app- Google Developers BlogIntroduced the Google Prediction API, a service that allows developers to easily integrate machine learning-powered predictive capabilities into their applications. This API enables apps to become 'smart' by leveraging user data to provide predictions, such as personalized recommendations and intelligent automation.
  66. Prediction API: Make smart apps even smarter- Google Developers BlogIntroduced the Prediction API, a new service that allows developers to integrate machine learning models into their applications for prediction tasks. This API provides a scalable and accessible way to leverage Google's machine learning infrastructure, enabling the creation of 'smarter' applications.