CEOs Say AI Is Making Work More Efficient. Employees Tell a Different Story.

Key Takeaway: Executives are bullish that AI is dramatically boosting productivity, but most employees report little-to-no time savings and say the real cost is stress, rework, and uncertainty about what AI can reliably do.

More Insights:

• Most nonmanagers say AI saves them under 2 hours a week—or none at all, while many executives claim it saves 8+ hours.

• Workers are more likely to feel anxious or overwhelmed by AI; the C-suite is more likely to feel excited.

• Many people use AI for simple tasks (search replacements, first drafts), not for high-value work (analysis, coding).

• A major drag is the “AI tax”: time lost to fixing errors, correcting hallucinations, and redoing work.

• Even CEOs admit the payoff is limited so far: only a small share report clear cost + revenue gains, and many see no significant financial benefit yet.

Why it matters: The biggest risk isn’t that AI “doesn’t work”—it’s that leadership measures potential while employees live the implementation, and that gap can turn a promising tool into a credibility crisis where “productivity gains” quietly become redistributed labor: more checking, more cleanup, and more pressure—without the time savings anyone promised.

Build voice-led AI, faster

Text was the first interface for AI. Voice is the next foundational layer.

If your product can’t listen, it’s missing a core connection to the real world. Gladia makes it easy to add production-ready speech-to-text so your AI can understand conversations, calls, meetings, and live interactions from day one.

With real-time and async transcription, 100+ languages, built-in diarization, and privacy-first controls, Gladia gives builders a reliable voice layer without months of infrastructure work.

Try the API in our playground and make voice a fundamental part of your AI stack in minutes.

Anthropic has to keep revising its technical interview test as Claude improves

Key Takeaway: Anthropic keeps redesigning its take-home technical interview because newer Claude models can now match or beat even top human candidates, making the test less effective at distinguishing real skill.

More Insights:

• Since 2024, Anthropic’s performance optimization team used a take-home test to screen for genuine expertise.

• As AI coding tools improved, candidates could increasingly “solve” the test by leaning heavily on Claude.

• Claude Opus 4 reportedly outperformed most applicants under the same time limit; Opus 4.5 then matched the best humans too.

• That created a measurement failure: the test started reflecting which model you used, not how good you are.

• The team responded by designing a newer, more novel test (less about hardware optimization) and even shared the old test publicly to invite better ideas. “If you can best Opus 4.5,” the post reads, “we’d love to hear from you.”

Why it matters: This isn’t just an “AI in schools” problem—it’s a warning that when tools become near-expert, traditional evaluations stop measuring capability and start measuring access and prompting, forcing institutions to rethink what “skill” even means in an AI-native world.

Humans& thinks coordination is the next frontier for AI, and they’re building a model to prove it

Key Takeaway: Humans&, a new startup founded by top alumni from major AI labs, is betting that “coordination” (not smarter chat) is the next frontier—and raised a massive seed round to build a socially intelligent foundation model plus its own collaboration platform.

More Insights:

• The company wants AI that can manage real collaboration: priorities, decisions, and long-running team alignment.

• They’re designing a new model architecture focused on “social intelligence,” not just Q&A or code.

• Humans& has no product yet, but hints it could replace multi-user tools like Slack/Docs/Notion by owning the collaboration layer.

• Training will lean on long-horizon reinforcement learning and multi-agent RL to plan, revise, and follow through over time.

• Big risk: it’s competing not only with productivity tools, but with AI giants already building collaboration features—and it must keep funding compute-heavy model training.

Why it matters: Coordination is where organizations actually win or lose—strategy, meetings, decisions, and follow-through—so if AI can reliably improve how groups think and act together, it won’t just automate tasks; it could reshape power inside companies by making “alignment and execution” a product, not a personality trait.

Other stuff

• The Messy Human Drama That Dealt a Blow to One of AI’s Hottest Startups

• Apple plans to make Siri an AI chatbot, report says

• eBay Explicitly Bans AI “Buy For Me” Agents, Updates Arbitration & Dispute Rules In User Agreement Update

• YouTube will soon let creators make Shorts with their own AI likeness

• Google DeepMind CEO is ‘surprised’ OpenAI is rushing forward with ads in ChatGPT

• Google Nabs Top Talent From AI Voice Startup Hume AI

• D4RT: Teaching AI to see the world in four dimensions

Hope you enjoyed today's newsletter

⚡️ Join over 300,000 people using the Superpower ChatGPT extension on Chrome and Firefox.