Google’s Gemini 2.5 Flash-Lite Launch: Google has launched Gemini 2.5 Flash-Lite, its fastest and most cost-efficient model in the  Gemini family. Priced at just $0.10 per 1M input tokens and $0.40 per 1M output  tokens, Flash-Lite delivers high performance at a fraction of the cost (Gemini 2.5 Pro, by comparison, costs $1.25/$10 per 1M tokens). According to Google’s announcement, the model’s “native reasoning” can be toggled on for complex tasks, and it maintains a 1 million-token context length and tool-use abilities like web search and code execution. In short, Google is pushing the frontier of “intelligence per dollar” with this release.

OpenAI’s Agentic ChatGPT: OpenAI rolled out a new “agent mode” in ChatGPT  that lets the AI act autonomously for complex, multi-step requests. Dubbed  ChatGPT Agent, it unifies the capabilities of OpenAI’s earlier tools (the web-browsing “Operator” and the analytical “deep research” modes). Users can now ask ChatGPT to perform tasks like scheduling events, purchasing items, or analyzing documents end-to-end. The agent can navigate websites, use APIs, run code, and even produce slide decks or spreadsheets autonomously. This feature is live for Plus/Pro users (via an “agent mode” toggle). OpenAI says this marks a major step toward AI that “can now actively engage websites—clicking, filtering, and gathering more precise results” all within one conversation.

GitLab’s AI DevOps Agents: Developer teams got a boost with  GitLab Duo Agent Platform entering public beta. GitLab’s new platform orchestrates multiple specialized AI agents across the DevSecOps lifecycle. For example, a Software Developer Agent can refactor code while a  Security Analyst Agent scans for vulnerabilities, all coordinated in parallel. The beta includes IDE extensions for VS Code and JetBrains, and will expand across GitLab’s products next month. GitLab’s goal is to offload routine dev tasks to AI agents, allowing devs to focus on creative work. (In related news, YugabyteDB  announced new features for AI developers – from integrated vector search to LlamaIndex support – signaling databases are also evolving to meet AI’s needs.)

Should you run AI models in the cloud or on your own devices? Local LLMs are gaining traction thanks to projects like llama.cpp and quantized model techniques that let you spin up sizeable models on consumer hardware.

Running a 13B-parameter model on a decent GPU is now feasible, offering low latency and full data privacy on-premises. Tools like  Ollama and  LMStudio  even make it easy to download and run open-source models (Llama 2, Llama 3, Mistral, etc.) on your laptop with minimal setup. This gives organizations control over sensitive data and avoids ongoing API costs.

On the other hand, cloud LLMs (think OpenAI or Google’s APIs) offer instant access to the latest, most powerful models and features – often months before open-source equivalents catch up. They scale on demand (no need to buy GPUs).

The bottom line: local LLM  deployments win on privacy, control, and possibly long-term cost for steady workloads, while cloud LLMs win on convenience, cutting-edge capabilities, and handling spiky or heavy workloads without hardware investment. Many teams opt for a hybrid approach, using local models for core private tasks and calling cloud APIs for overflow or specialized tasks.

Thinking Machines nabs $2B: In a jaw-dropping early-stage round, Thinking Machines Lab – the AI startup founded by former OpenAI CTO Mira Murati – raised $2 billion at a $12 billion valuation. The  Andreessen Horowitz-led round (with participation from Nvidia, AMD, Cisco, ServiceNow, and others). Murati says the first product, due in a few months, will have an open-source component and aims to make AI “safer, more reliable” for a broad range of applications. The outsized funding underscores how coveted top AI talent has become amid the 2025 AI funding frenzy.

Musk’s xAI valued at $113B: Elon Musk’s new AI venture xAI is reportedly raising a $5 billion round that includes a $2B investment from SpaceX. The funding (alongside a large debt package) comes on the heels of xAI’s merger with Twitter (rebranded X) and the debut of its “Grok” chatbot. According to reports, the combined X + xAI company would be valued at $113 billion post-investment. Musk has hinted he’d welcome an investment from Tesla next – subject to board approval. xAI is spending heavily on training an advanced conversational AI to compete with OpenAI, even as Grok’s edgy style has courted controversy.

Composio’s $29M for AI that learns: San Francisco-based  Composio raised $25M (total $29M with prior seed) to build AI agents that learn from experience. Lightspeed Venture Partners led the Series A, joined by tech luminaries like Guillermo Rauch (Vercel) and Dharmesh Shah (HubSpot). Composio’s platform provides a shared “learning layer” for AI agents – so an agent that figures out one task can transfer that knowledge to others. Why it matters: Today’s LLM-based agents “don’t get better at their jobs the way a human would”, CEO Soham Ganatra explains. They hit a wall of forgetting. Composio’s infrastructure aims to change that, letting agents accumulate skills over time. The funding will accelerate the development of this “memory” layer for agents in enterprise workflows.

Enterprise AI funding surge: Compliance and security automation startup  Vanta secured $150M (Series D) at a $4.15B valuation, a 69% jump from last year. The round, led by Wellington Management, will help Vanta inject more AI into its cloud compliance platform (which helps companies meet SOC2, ISO27001, etc.). Meanwhile, the White House’s focus on AI has spurred defense investments too – e.g., Shield AI (military autonomous drones) reportedly raised $200M in July. And in the Middle East, Saudi Arabia’s new state-backed AI company  Humain announced a $10B venture fund  to invest in global AI startups, part of a plan to capture “7% of global AI workloads” with massive data centers and strategic partnerships.

Elkar – Your Spreadsheet Copilot: Also on our radar is  Elkar, an AI analyst that lives inside Excel and Google Sheets. If you’re dealing with complex spreadsheets, Elkar is a game-changer. It’s basically  AutoGPT for Excel – you describe what you need in natural language, and Elkar will do it. It can write formulas, build charts and dashboards, clean up messy data, and even debug errors in your sheet. For example, say you have sales data with inconsistent date formats and some missing values: just prompt Elkar to clean the data and highlight outliers, and watch it work magic. It’s like having a junior data analyst who never sleeps embedded in your spreadsheet. Elkar was launched recently (and even got a nod from Microsoft’s AppSource). We found it especially useful for generating complex Excel formulas without Googling or remembering functions – “SUMPRODUCT with criteria”? Just ask Elkar. If you’re in finance, analytics, or just Excel-heavy roles, keep an eye on this tool for a productivity boost.

White House AI Action Plan: The U.S. administration unveiled “Winning the AI Race: America’s AI Action Plan,” a sweeping federal blueprint for AI development and governance. Announced on July 23 at a D.C. summit, the plan calls for  90+ actions to accelerate domestic AI innovation, remove regulatory barriers, and promote AI infrastructure. It emphasizes boosting R&D investments and even loosening certain rules (like easing environmental and export restrictions). At the same time, the plan highlights  Responsible AI  principles: developing AI systems that are transparent, reliable, and safe, with evaluation tools to ensure models meet standards for accountability and factual accuracy. The White House is pressing for industry partnerships to shape standards on issues like watermarking AI content and mitigating bias. In short, the U.S. is adopting a “light-touch, pro-innovation” stance – trying to speed up AI progress while self-policing risks via guidelines and voluntary commitments. (This contrasts with the EU’s heavier regulatory approach.)

Corporate AI governance: In industry news, OpenAI published a transparency report and system card for the new ChatGPT agent, detailing safeguards against autonomous misuse. And  Google released its latest AI  ESG report, highlighting energy efficiency gains in its data centers and new tools for identifying bias in training data. Many AI companies are proactively disclosing these measures as regulators globally inch toward formal AI oversight.

How do you keep an AI model up-to-date and grounded in facts? Enter  Retrieval-Augmented Generation (RAG). RAG is a technique to enhance a generative model’s accuracy and reliability by feeding it relevant external information. In practice, a RAG system will search a knowledge source (documents, database, the web) for content related to your prompt, and augment the model’s input with those results.

The model then generates its answer with that extra context, which helps reduce hallucinations and improve factual correctness. For example, if you ask a question about current events (beyond the AI’s training data cutoff). The RAG approach typically involves four steps: ingestion of authoritative data into a retrievable index, retrieval of the most relevant chunks, augmentation (merging those chunks with the query), and generation of the response using this augmented prompt. This way, the model’s knowledge isn’t limited to its frozen training data. RAG is increasingly popular in enterprise AI apps, since it allows the use of private data and sources for custom Q&A, and the model can also cite sources (building user trust in its answers). In summary, RAG gives AI a kind of open-book exam – letting it fetch facts as needed – rather than relying purely on parametric memory.

This week’s prompt idea comes from a great post we saw on leveraging ChatGPT for entrepreneurship.

This prompt guides AI to generate startup or product ideas tailored to the user, and by asking for trends, it ensures the suggestions are grounded in some market reality. We tried it and got some surprisingly insightful ideas (one in sustainable fashion rental, and another in AI-driven personal fitness coaching, complete with trend citations about eco-conscious Gen Z and post-pandemic health focus!). The key is to be specific about your own skills or domain in the prompt. Pro tip: after the first output, ask follow-ups like “Great, now give me a lean 3-month execution plan for idea #2”. Using a structured prompt like this can turn ChatGPT into a brainstorming partner that not only tosses out ideas but also backs them up with reasoning. Happy prompting – may one of these suggestions spark your next side hustle!