Search the archive

build agentic AI systems. Build with Agent Workbench

ENABLING TRILLION-PARAMETER MODELS ON AWS EFA

Perplexity published code enabling trillion-parameter mixture-of-experts models to run efficiently across multiple AWS servers for the first time, opening up less advanced data centers to AI inference. The challenge is that large models, even many open weights models, are too large to fit on single servers, but AWS' networking hardware was too slow for the constant inter-server communication these models require.

STRUCTURED OUTPUT IN GEMINI API

Google enhanced the Gemini API with expanded JSON Schema support and improved property ordering, making structured outputs more reliable for use in multi-agent workflows and database tasks.

PINTEREST CEO TOUTS OPEN SOURCE AI: ‘TREMENDOUS PERFORMANCE' WITH REDUCED COSTS

Pinterest plans to move forward with open source models for its various use cases. They come at a fraction of the cost of the larger model providers, and they deliver value to users effectively. The company will be observing user feedback to see if people actually want AI shopping agents before implementing them. The company is planning to roll out personalized boards curated with AI.

SNAP AND PERPLEXITY SIGN $400 MILLION DEAL TO PUT AI SEARCH DIRECTLY IN SNAPCHAT

Perplexity, an AI search engine, will appear directly in Snapchat sometime in early 2026. It will become a prominent part of Snapchat's 'chat' interface. Perplexity is paying Snap $400 million for the integration. Snap may pursue similar partnerships with other AI companies.

no-code backend that isn't a black box. Explore

OPENAI CALLS FOR $1T INVESTMENT TO SECURE US AI LEADERSHIP

The US risks losing its AI leadership due to energy shortages, threatening economic growth and technological advances.

HOW TWO SILICON VALLEY A-LISTERS BECAME BILLIONAIRES BY REMAKING CUSTOMER SERVICE WITH AI

Bret Taylor and Clay Bavor started Sierra, a $10 billion startup that builds AI customer service agents, in 2023.

A SHORT LESSON IN SIMPLER PROMPTS

You can build much shorter, more effective prompts by working with AI models instead of fighting them and balancing AI with plain old engineering.

Explore the full report >>

ANTHROPIC PROJECTS $70B IN REVENUE BY 2028

The company projects API revenue of $3.8 billion this year, double OpenAI, and is targeting $20-26 billion in annual recurring revenue for 2026. OpenAI is aiming for $100 billion in revenue by 2027, but expects to lose $115 billion before reaching profitability in 2029, compared to Anthropic, which is projected to be cash flow positive by 2028.

GOOGLE IS PREPARING NANO BANANA 2 FOR THE UPCOMING RELEASE

Google's upcoming GEMPIX2 model (Nano Banana 2) could be available as soon as next week. It is expected to be targeted at creators and professionals who rely on Gemini for image generation and AI-assisted creative workflows. The Nano Banana line has attracted considerable interest and contributed to a noticeable growth in Gemini's user base.

OPENAI'S SORA VIDEO GENERATOR APP FINALLY AVAILABLE ON ANDROID

Sora is now available on Android, making hyperrealistic video generation capabilities available to millions more users. The app quickly became a social media phenomenon when it was first released on iOS. It allows users to create realistic videos with audio and impressive physics. Android users will still need an invite code to actively use Sora's generation features. The launch is limited to select countries, including the US, Canada, Japan, Korea, Taiwan, Thailand, and Vietnam.

CHAINING FFMPEG WITH A BROWSER AGENT

100x included ffmpeg into its Chrome extension to make complex media processing a single, serverless, and stateless step in any workflow. It already had a browser agent for automation - ffmpeg became another 'tool call' for the agent. This enabled the agent to deal with complex editing tasks like mixing audio and video and adding text over video. This post gives readers a look at how the integration works.

BEYOND STANDARD LLMS

The largest and most capable open-weight large language models today are autoregressive decoder-style transformers. Some alternatives have popped up in recent years. While some are geared toward better efficiency, others are aimed at improving model performance. This article gives an overview of alternative models that exist, from text diffusion models to the most recent linear attention hybrid architectures.

HOW WE BUILT A CUSTOM VISION LLM TO IMPROVE DOCUMENT PROCESSING AT GRAB

Grab developed a lightweight, specialized Vision LLM from the ground up to address the limitations of traditional Optical Character Recognition (OCR) systems. Traditional OCR systems struggle with the variety of document templates they have to process and often fall short in understanding South East Asian languages. Grab's model can effectively replace traditional OCR pipelines, opening new possibilities for document processing at scale. This post details how the company built its new model and also provides performance results.

CODE EXECUTION WITH MCP: BUILDING MORE EFFICIENT AGENTS

Code execution can enable agents to interact with MCP servers more efficiently. When too many servers are connected, tool definitions and results can consume excessive tokens, reducing agent efficiency. Many of the problems with MCP feel novel, but they have known solutions from software engineering. Code execution applies established patterns to agents, letting them use familiar programming constructs to interact with MCP servers more efficiently.

RUN ANY LLM WITH A SINGLE API: INTRODUCING ANY-LLM V1.0

any-llm is an interface that can use any model, cloud or local, without requiring a rewrite of the stack every time. It makes it easy for developers to use any large language model without being locked into a single provider. any-llm decouples project logic from model provider with a stable and consistent API surface.

PROFILING WITH CURSOR 2.0: THE MISSING LAYER IN AI CODE GENERATION

Profiling is the best way to separate code that looks good from code that is actually good. The Pyroscope Performance Profiler is a Cursor extension that takes profiling data and paints it right on the code. It allows developers to look at multiple solutions generated by different agents and pick the best based on real numbers. Performance comparisons across deployments, trend analysis, and incident inspection can all be done inside the editor without going to a separate UI.

EXPLORING A SPACE-BASED, SCALABLE AI INFRASTRUCTURE SYSTEM DESIGN

Google will launch two satellites carrying its AI chips, called TPUs, in 2027, in hopes that in the mid-2030s space-based data centers will become more cost efficient since solar panels are up to 8x effective in orbit. Project Suncatcher envisions a constellation connected by optical inter-satellite links to power distributed ML tasks, like AI inference.

AMAZON SENDS PERPLEXITY A CEASE AND DESIST OVER ITS AI AGENTS SHOPPING FOR YOU

Amazon has demanded that Perplexity stop its agentic web browser from buying products on behalf of humans. It claims that the Comet browser degrades the shopping and customer service experience that Amazon provides. Amazon has accused Perplexity of violating its terms of service and committing computer fraud by failing to disclose when an AI is shopping for a user. Perplexity responded by accusing Amazon of bullying and using litigation to stifle innovation.

THE COMPANY QUIETLY FUNNELING PAYWALLED ARTICLES TO AI DEVELOPERS

At the heart of AI web scraping is Common Crawl, a small nonprofit that supplies, and is funded by, major AI labs with internet data while flagrantly disregarding paywalls and removal requests. The organization has gathered millions of articles from The New York Times, The Atlantic, and other publishers (many of them now suing AI companies), but has avoided compliance by making its search tool return "no captures" for domains that actually contain thousands of articles.

bridges the gap

GEMINI'S RETENTION DATA IS OUT, AND IT LOOKS BETTER THAN EVEN CHATGPT!

Three-month retention is consistently over 90%.

CODING AGENTS ARE OUTLIERS

The success of coding agents has created unrealistic expectations for AI agents in other domains.

SHOPIFY SAYS AI TRAFFIC IS UP 7X SINCE JANUARY, AI-DRIVEN ORDERS ARE UP 11X

Shopify is working with Perplexity and Microsoft to add other in-chat shopping experiences after the success of its integration with ChatGPT.

CODEMAPS: UNDERSTAND CODE, BEFORE YOU VIBE IT

Cognition launched Windsurf Codemaps, which creates AI-annotated structured maps of codebases to address the onboarding costs of learning a new codebase or maintaining unfamiliar legacy code.

COMMITMENTS ON MODEL DEPRECATION AND PRESERVATION

Anthropic will preserve weights of all publicly released Claude models for the company's lifetime and conduct exit interviews with models before retirement, documenting their preferences about future development.

TRY WISPR FLOW FOR FREE →

AWS AND OPENAI ANNOUNCE MULTI-YEAR STRATEGIC PARTNERSHIP

OpenAI is moving a significant portion of its compute to AWS in a $38 billion, seven-year deal that provides immediate access to hundreds of thousands of Nvidia GPUs with the ability to scale to tens of millions of CPUs. OpenAI won't serve its models through AWS (yet) due to its exclusive cloud relationship with Microsoft.

MICROSOFT SIGNS $9.7 BILLION CLOUD DEAL WITH IREN AS AI DEMAND SWELLS

Microsoft inked a $9.7 billion deal with data-center operator IREN to access Nvidia's advanced chips, addressing a computing crunch amid increasing AI demand. This five-year partnership allows Microsoft to boost capacity without constructing new facilities or securing additional power. IREN's phased deployment, supported by Microsoft's prepayment, will run through 2026.

WHAT'S UP WITH ANTHROPIC PREDICTING AGI BY EARLY 2027?

To get Nobel-Prize-winning AIs by early 2027, AIs must fully automate researching engineering by December 2026 and go from completing hour-long tasks today to completing two-week-long tasks in just the next 8 months. OpenAI is only slightly more conservative than Anthropic, estimating automated AI research by March 2028.

BABY SHOGGOTH IS LISTENING

AI will replace human writers and readers. Influential figures in AI like Tyler Cowen and Gwern argue that writers shouldn't optimize for human readers served by algorithmic feeds but for LLMs that will invariably train on their work. The motivation ranges from practical (LLMs now mediate most information retrieval and composition) to existential (influencing a future superintelligence may be the only lasting legacy).

AI-driven workflows. Save your spot now 💺

QWEN3-MAX-THINKING

Qwen3-Max-Thinking has been released in early preview. It currently achieves 100% on challenging reasoning benchmarks like AIME 2025 and HMMT when augmented with tool use and scaled test-time compute. The model is available in Qwen Chat and the Alibaba Cloud API.

NOTES FROM READING THE AMP MANUAL

Amp is a coding agent with support for MCP, the ability to set project, folder, and global context, a permissions system, headless mode, and more. It stores all threads on the server so users can resume a thread anytime. Amp picks the right model for each task, but it can be nudged to use more expensive models. It is pay-per-token and not cheap, but it is worth it because it saves time.

ASSISTANT-UI

assistant-ui is an open-source TypeScript/React library for building production-grade AI chat experiences quickly. It handles streaming, auto-scrolling, accessibility, and real-time updates, and works with any custom backend. assistant-ui has broad model support out of the box. It is fast to production and designed for customization with a great developer experience.

ANTHROPIC LANDS COGNIZANT AS ENTERPRISE AI CUSTOMER

Cognizant Technology Solutions is now one of Anthropic's three largest customers by total number of enterprise seats. The company plans to roll out Claude to its 350,000 employees. It will also co-sell Anthropic's Claude models to its business customers. The financial terms behind the partnership have not been publicized.

PERPLEXITY ASSISTANT WILL BE ABLE TO JOIN YOUR MEETINGS SOON

Perplexity is developing a wider set of meeting automation tools. These upcoming options will let users toggle automatic recordings of meetings, enable the assistant to join certain meetings on its own, and specify recipients who should receive a recap after the session concludes. They will help professionals and teams streamline meeting documentation and follow-up and cut down on manual note-taking and coordination. There is no timeline announced yet for the rollout of these new features.

and 25 minutes of Q&A. Register for all three sessions, free

THIS IS THE LETTER THAT CAUSED GEMMA TO BE PULLED FROM AI STUDIO

Republican Senator Marsha Blackburn wrote a letter to Google demanding information about why its small, open-source model Gemma hallucinated “defamatory” information about her.

MICROSOFT TO INVEST OVER $15 BILLION IN UAE, SECURES US EXPORT LICENSES FOR NVIDIA CHIPS

The investment primarily funds data center expansion, where cutting-edge Nvidia chips will be allowed to enter the UAE for the first time.