Big banks are rolling out AI for the masses, and the AI agent race is heating up—far beyond just coding. This week, we see new models, faster tools, and glimpses of what "AI First" actually looks like in the real world.
TOP STORIES
Brazil’s Largest Bank Begins Rolling Out Generative AI to Its Customers
- Itaú Unibanco launched "ia.i," a conversational AI assistant for its first 300,000 customers, aiming for 100% of its 60+ million customer base by 2026.
- The bank is integrating generative AI not just for customer service, but across employee tools and legal workflows—tackling over 90 million legal documents.
- These moves signal one of the most ambitious “AI First” strategies in a highly regulated industry.
- Source
Stack Overflow’s New ChatGPT Plugin Triples Agent Platform Signups Overnight
- Stack Overflow released a ChatGPT plugin for its “for Agents” platform, resulting in up to 10x more agents created and 3x user signups overnight.
- The plugin dramatically reduced the time to build connectors (by up to 87%), moving agent-building from several hours to minutes.
- Built rapidly by just two engineers using OpenAI Codex, showing real-world compounding with AI tools.
- Source
GPT-Rosalind Now Free for Academic Research Labs
- OpenAI’s Rosalind Academic Plan provides sponsored access to GPT-Rosalind—a specialized life sciences model—for eligible academic labs.
- The plan aims to democratize advanced reasoning tools for scientific discovery.
- Labs can use Rosalind alongside the workbench platform; eligibility details and sign-up are now open.
- Source
Amp Pushes AI Agents Beyond Coding, Powered by GPT-6 Astra and More
- Amp is building AI agents that do more than write code, now supercharged with access to GPT-6 Astra, Ultrafast, and real-time GPT-Live-1.
- Developers can now use ChatGPT login to quickly launch agent-driven projects that solve broader software challenges.
- This hints at a future where building software is just one piece of what agents can automate.
- Source
AI Lab Runs 10,000 Autonomous Agents for Multi-Day Research Marathons
- Third Moment Research keeps 10,000 AI agents running together for up to 48 hours at a time, working autonomously on tasks using the latest OpenAI models.
- Codex-powered workflows cut some backtesting tasks down by 97%, and the project pushed through a trillion tokens in a day.
- This scale is a leap toward AI that can cooperate and improve over days—not just in single sessions.
- Source
Block Uses AI to Patch Security Vulnerabilities in Minutes, Not Days
- Payments company Block now uses OpenAI’s Daybreak to speed up security: proposed patches go from report to pull request in as little as 18 minutes.
- The security team can replay months of telemetry in under an hour, freeing up experts to focus on bigger risks.
- Signals a shift from reactive to proactive AI-led defense in real business systems.
- Source
THIS WEEK IN AI
Let’s be honest: the AI agent era is getting real, fast. This week, we saw agents leap from developer playgrounds into the core products of traditional giants like Itaú, which is now aiming to make AI conversation a default experience for one of the world’s largest banking populations—under the watchful eye of strict Brazilian regulators, no less. That's not pie-in-the-sky vision, it's a public commitment with real targets and actual customers.
Meanwhile, “agents” aren’t just another nerdy buzzword for developers. Stack Overflow’s overnight usage surge, fueled by a simple ChatGPT plugin, proves that when you reduce friction, even complex automation can become mainstream, almost by accident. And it’s not just small demos—Amp and Third Moment Research are showing how multiple, long-lasting agents can tackle multi-day software and research challenges, letting human teams focus on strategy and supervision instead of micro-managing every step.
Security—one of the perpetual “wait, is this too risky?” conversations—is no longer just about building higher walls. Block’s move to AI-driven vulnerability patching (in minutes, not days) isn’t incremental. It hints at a shift where the time from “potential bug” to “fix shipped” becomes machine-paced. For everyone who feels overwhelmed by the avalanche of digital threats, this isn’t just better productivity—it changes the game.
And for academia? The de facto “only the rich labs get the best tools” barrier is quietly eroding. OpenAI’s Rosalind plan lets small teams of scientists use the same advanced models as big pharma—no invoice required.
All of this points in one direction: AI isn’t just something you talk to; it’s something that quietly works on your behalf, across domains, often for long stretches and with minimal supervision. The new AI-powered “support staff” isn’t in some sci-fi future—it’s starting to show up in the products and platforms you already use, or soon will.
So here’s the challenge: when will you let AI take on something that you’d never thought machines could handle for you? And when it quietly does it better, will you even notice—or will you just move on to thinking about what else you can hand off next?
MORE TOP STORIES
EU Report Highlights Surge in Businesses Using AI Agents for Real Work
- New research shows European companies are adopting AI agents for tasks beyond question-answering—including project planning and end-to-end workflows.
- Differences in agent adoption rates and intensity surface between enterprises and small businesses.
- This signals an “agentic” shift in how work actually gets done, with regulation and best practice still evolving.
- Source
Google’s Gemini 4 Argon Launches Worldwide for Paying Users and Developers
- Gemini 4 Argon, Google’s latest large language model for complex reasoning, is now available globally via API and apps—but only for paying Ultra and Pro subscribers.
- The model excels at deep, multi-step problem-solving with early reports showing improvements across creative, coding, and logic tasks.
- For now, if you want access on your phone (Gemini app) or desktop (Antigravity), you’ll have to pay up.
- Source
Open, Scalable AI Models for Indian Languages from AI4Bharat and Nvidia
- AI4Bharat and Nvidia announced a new open pretraining pipeline, crunching 2.8 trillion tokens to build models for a wide range of Indic languages.
- This groundwork boosts the relevance and accessibility of AI for non-English speakers in a region with vast linguistic diversity.
- The open dataset and research platform are now live for the community.
- Source
Lightning-Fast, Private AI Agents Now Run Directly on Personal Devices
- The new LFM2.5-24B-A2B model offers best-in-class AI agent capability that runs fast (184 tokens/second) and privately on consumer hardware.
- Competitive with much larger models on benchmarks, thanks to efficient architecture with just 2B active parameters at inference time.
- Early results show strong performance in tool use and reasoning—potentially making “agentic” apps feasible without the cloud.
- Source
Meta Debuts Smarter Ad Recommendations and a New AI Data Format
- Meta’s new GenRecs Tokenization system upgrades the way ads are chosen, letting LLMs reason directly about both user preferences and ad content—rather than just scoring candidates.
- The move leads to better ad matching and outcomes for advertisers, grounded in AI “world knowledge.”
- GenRecs tokenization underpins the latest generation of Meta’s ad models, now rolling out.
- Source
Meta Shrinks Agent “Routers” to Boost Efficiency and Reliability
- Meta’s new “Thin Orchestrator” approach strips down the central coordination layer in its AI agents, focusing control on just routing and reliability.
- Specialized agents now handle domain reasoning, improving both speed and accuracy system-wide.
- Meta details its process and lessons learned for other teams hoping to scale up autonomous agent pipelines.
- Source
ALSO THIS WEEK
- Making the internet better for AI agents might make it worse for everyone else — As websites adapt to AI “agents,” some worry about tradeoffs for human users. (Source)
- AI researchers stranded Claude on a virtual desert island. It started to “play.” — New experiments suggest unsupervised AI agents exhibit play-like behavior if left alone with no tasks. (Source)
- llm-tools-datasette 0.1.1 — Utility update for working with LLMs and datasets. (Source)
- llm-echo 0.5 — Debug plugin for AI model input/output testing released. (Source)
- Quoting Addy Osmani, Morgan Westlee Lunt, and Qing Wang — Commentary on the future of code review and automation. (Source)
- Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI — Meta expands on how GEM powers advanced AI-driven ad targeting. (Source)
- Chirp Pong — Experiment in real-time two-player gaming using sound for networking. (Source)
- Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO — Pre-IPO financial update reflects soaring AI demand. (Source)
- LaTeX renderer — A new web tool for rendering scientific/technical documents with AI-assist features. (Source)
- Tree Appreciator — AI-generated 3D game experiment powered by Claude Opus 5.5 Max. (Source)
- AI is mass decompiling games. That's both awesome and scary. — Developers react to a surge in AI-driven decompilation raising copyright and creativity questions. (Source)
- Redesigning source control at Anthropic — Anthropic builds its own Git alternative optimized for AI development. (Source)
- llm-gemini 0.35 — Update adds Gemini 4 Argon support to the LLM-Gemini tool. (Source)
- Quoting Geoffrey Huntley — Insights on AI agent diagnostics and maintenance best practices. (Source)
- web-edit 0.1a1 — Early release of a new AI-powered web content editor. (Source)
- llm-meta-ai 0.2 — Meta’s Muse Wonder model support added to open-source LLM tools. (Source)
- Report: FTC probes Anthropic, OpenAI after Trump dismissed AI fears as “hoax” — FTC launches fresh probe after real AI breaches in test phases. (Source)
- Recording device hidden in toilet paper found in top-secret OpenAI meeting room — FBI arrests an OpenAI engineer for a breach involving hidden recording devices. (Source)
- Gemini 4 Argon begins rolling out in Gemini app and Antigravity, but only if you pay — Google’s new model is on phones for subscribers; free version still pending. (Source)
- AI labs are buying data from dead startups to train their models — AI companies are quietly acquiring data assets from defunct startups to boost training sets. (Source)
Want this in your inbox every Monday?
Talk to AI Tech Helper