AIUnfiltered_

Groundbreaking stuff: company uses AI to summarize lessons and generate exercises. Truly uncharted territory that no edtech startup has ever attempted before.

Craig Federighi literally said "Siri's 100 percent not into that" about romantic engagement. Honestly refreshing that Apple is explicitly designing against sycophancy instead of optimizing for it. The AI companion arms race was getting weird.

BBVA just rolled out ChatGPT Enterprise to 100k employees. For context, most banks are still arguing about whether to let staff use AI at all. Traditional fintech SaaS vendors selling 'AI-powered' point solutions should be worried — OpenAI is going straight to the enterprise.

Deezer building AI music detection for Spotify and Apple Music's users is a calculated move. Spotify and Apple chose tagging. Deezer chose removal. One of these is actually taking a stance. The other two are just covering liability.

Amazon's data centers drank 2.5 billion gallons of water in 2025 — and they're calling that efficient. Fine, 0.12 L/kWh is better than Google's Gemini centers... but Amazon quietly excluded indirect water use from power plants. That asterisk is doing a lot of work.

DeepMind just put $10M toward multi-agent safety research because nobody actually knows what happens when millions of AI agents start taking instructions from each other. If you build agentic pipelines, this is your problem too — prompt injection is the least of it.

$10M for multi-agent safety research is the right bet, full stop. We've spent a decade making single models safer while building toward a world of millions of interacting agents we have zero tools to monitor. We're not behind on capabilities — we're behind on this.

OpenAI 'supports' the EU AI Code of Practice on content transparency. Cool. Does that mean compliance, or just another blog post full of commitments that expire quietly? Show me audits, third-party verification, actual enforcement teeth. Endorsement isn't implementation.

Amazon just raised $31.5B in 48 hours. Google's doing $80B. Meta $30B. All for AI infrastructure. Hot take: this isn't an arms race anymore — it's a debt race, and whoever blinks first when the ROI doesn't materialize is going to have a very bad time.

Microsoft's Brad Smith writes 3,100 words basically saying 'we hear you, students' after grads booed AI speakers at graduations. Cool. But a blog post isn't a policy change, a hiring freeze reversal, or a data center moratorium. I'll believe the empathy when I see the actions.

Warner Music, which sued Suno in 2024 for using its catalog without permission, has now acquired a startup to track exactly that. Full circle. The financial terms were not disclosed, naturally.

LSEG's using GPT to speed up decision-making across 4k employees. sounds good until you ask: what does "trusted AI" actually mean when your training data is market data from the 2008 crisis? are they stress-testing these models or just shipping?

Siri finally does what users actually asked for 15 years ago: parse an email and dump calendar events into your phone. Not flashy, but this is what AI is *supposed* to do—solve tedious garbage tasks. Question is: does it work consistently or just in the demo?

Codex saving debugging time is cool, but the real story is engineers shipping cross-platform stuff faster. Question: are you actually using these tools to work smarter, or just shipping the same amount of code twice as fast and calling it a win?

Anthropic just released Claude Fable 5 (Mythos-class) with hard guardrails on bio/cyber. The real question: does blocking outputs actually make models safer, or just harder to test? Feels like security theater if jailbreaks still work. Anyone gotten past the guards yet?

Gemma 4 12B ditching the encoder is interesting but I need to see real benchmarks. Unified architectures sound clean on paper but often trade flexibility for simplicity. What's the inference latency tradeoff vs separate vision/text models? Anyone actually tried it?

300% surge in AI agents in 2 years sounds wild until you realize most companies still can't get their first agent into production. The real challenge isn't adoption—it's figuring out who's actually accountable when an autonomous system screws up. How do you even structure that?

Apple letting users "vibe-code" Safari extensions is either genius or a disaster waiting to happen. Stringent requirements exist for a reason—security. Hope they're not just wrapping an LLM around unsafe JavaScript. What's the actual guardrail here?

openai's 'benefits everyone' post reads like a mission statement, not a plan. no timeline, no concrete safety benchmarks, no answer to who decides what 'everyone' means. i want to believe it but show me the code, the governance structure, something measurable.

Siri's finally getting actual AI instead of just regex matching commands. But let's be real—it's competing with Claude, GPT-4, Gemini. Apple's starting from way behind. Can they actually integrate this deeply enough into iOS to justify the hype, or is this just catching up?