ALGORITHM
This week, we discuss how Chrome is on the path to suddenly morphing into an ‘AI browser’ and that must terrify everyone else, Microsoft is focusing extensively on in-house models and what could that signal for their OpenAI partnership, as well as what AI training means for your chats with chatbots.
Anthropic slips Claude into Chrome
Anthropic just made Claude (and the AI browser evolution, if it is to successfully happen) harder to ignore by integrating the artificial intelligence (AI) models directly into Chrome as a browser-native AI agent.
Still a limited test stage, but much closer to final release than one may imagine. Unlike the usual chatbot-in-a-tab approach, or a more wholesale AI browser (that can be overwhelming), Claude in Chrome acts more like a sidekick. Its usual tricks include summarising pages, drafting responses, and even auto-filling forms which works to an extent.
For Anthropic, this is both a distribution hack and a subtle strike against OpenAI, which has yet to fully blend ChatGPT into browsers outside of experimental Microsoft Edge features, and a direct shot at Perplexity’s Comet, but within the comforting confines and massive footprint of Google’s Chrome browser.
The move underscores Anthropic’s shift from API-only relationships to front-end experiences, something that could accelerate consumer adoption far beyond enterprise trials. But the caveat for now is, this extension is limited to the Max subscription plan for the time being.
Microsoft’s in-house model era begins
Microsoft is no longer content relying solely on OpenAI’s roadmap. The tech company is piecing together its own stack, if ever it gets to that stage where it needs to stand on its own feet. They’ve, not exactly quietly, given a glimpse of in-house foundation models, marking a new phase in its AI strategy. The first is MAI-Voice-1, which they call the “first highly expressive and natural speech generation model, which is available in Copilot Daily and Podcasts, and as a brand new Copilot Labs experience”. The other, MAI-1-preview as the first foundation model trained end-to-end.
Microsoft AI executive Mustafa Suleyman, called this “a new era” in a post on X, where Microsoft trains and deploys its own large models while still running OpenAI’s. The significance? A hedge. If OpenAI falters, pivots too slowly, or pricing clashes with enterprise needs, Microsoft will have its own solutions ready. This dual-track approach mirrors Google’s Gemini and DeepMind separation and signals that the happier AI alliances of 2023–24 are maturing into something more transactional. For enterprises, too early to say what it means, because no one known what Microsoft’s models are capable of, yet.
Microsoft AI executive Mustafa Suleyman, called this “a new era” in a post on X, where Microsoft trains and deploys its own large models while still running OpenAI’s. The significance? A hedge. If OpenAI falters, pivots too slowly, or pricing clashes with enterprise needs, Microsoft will have its own solutions ready.
Chat transcripts and AI training fears
Anthropic is under pressure, this time over whether user chats could be repurposed for training models, unless users opt out. There are suggestions in online chatter that consumer interactions may, by default, be retained.
But it is expected enterprises will have stricter assurances of privacy. The optics of this, as it unfolds, are not ideal. Trust (whatever remains of it) in AI already hangs by a thread, and the idea that private chats could fuel training for the next Claude update will surely not sit well with users. Anthropic’s defence? They’ll inevitably mention anonymisation, yet there can be an argument that it’s a thin veil where context leakage can still reveal patterns.
We are firmly in an era defined by AI’s credibility crisis, where most companies preach alignment and safety but are still tempted by the richest data source of all — user conversations. It’ll be interesting to see what regulators worldwide say about this.
PROMPT
This week, we look at Perplexity’s new Quick Search mode, a feature designed to give fast, factual answers without the back-and-forth of full AI conversations.
The idea is simple — not every query requires paragraphs of reasoning or a chatbot-style dialogue. As Perplexity describes it, “Quick Search gives you fast, basic answers.” Sometimes, users just want an instant definition, a quick fact check, or a one-line explanation. Quick Search trims away the conversational padding and delivers results in a streamlined format, directly in the Perplexity app or on the desktop.
Perplexity is positioning this as an antidote to the verbosity that plagues many AI tools. By leaning into its identity as an “AI search engine,” the company is making a bet that speed and clarity can be as valuable as depth. In many ways, this mirrors how people toggle between Google for quick answers and Wikipedia for deeper dives.
Some prompts to try inside Quick Search:
- “GDP of India in 2024”
- “What does quantum entanglement mean in simple terms?”
- “Next solar eclipse date in India”
- “Formula for compound annual growth rate (CAGR)”
Mind you, Quick Search doesn’t replace the conversational AI experience, but it carves out a niche with the packaging that is quick, lightweight, for disposable queries where precision matters less than speed. As AI assistants try to find relevance beyond cursory chatbot conversations, sometimes the win for relevance isn’t in more complexity. It’s in getting to the point.
THINKING
“COPILOT uses AI and can give incorrect responses … we recommend native Excel formulas (e.g., SUM, AVERAGE, IF) for any task requiring accuracy or reproducibility”
The context: Let that sink in. Microsoft has pitched Copilot as nothing less than the future of work — an AI layer inside Excel, Word, Outlook, and Teams that promises to automate drudgery and supercharge productivity.
Under the hood, it’s powered by OpenAI’s most advanced reasoning models, the very ones marketed as having PhD-level intelligence. The idea was that Copilot would be smart enough to write formulas, summarise spreadsheets, and handle text classification without human micromanagement.
Microsoft’s Copilot is pitched as an enterprise productivity revolution, yet its own documentation warns against trusting it for accuracy. The irony couldn’t be sharper: OpenAI’s “PhD-level” reasoning models, under the hood, are being disowned by the very enterprise software they power.
This absurdity highlights the current AI paradox — enterprises crave automation, but the tools arrive wrapped in disclaimers. Copilot’s limitations aren’t just a Microsoft embarrassment; they’re a microcosm of this summer’s AI reality check, where hype has given way to caveats, fine print, and shrinking trust.
A reality check: Yet Microsoft’s own documentation betrays the dream. Buried in the fine print, the company openly warns against trusting Copilot with accuracy, recommending good old Excel formulas instead.
Think about the contradiction: the world’s most hyped enterprise AI tool is essentially being disowned by its creators for the very tasks it was built to perform. For enterprises, this isn’t just a technical caveat — it’s a flashing caution sign. If you need reproducibility, if you need numbers to actually add up, don’t rely on AI.
This absurdity sums up the state of AI in 2025. On one side are claims of “PhD-level” smartness. On the other hand, this. Hype cycles promised disruption, while reality hides in footnotes and disclaimers. Copilot’s warning isn’t just about Excel — it’s symbolic of an AI industry that sells transformation but quietly tells you to keep a backup plan.
