Hi Reader,

Here are three things I found interesting in the world of AI in the last week:

OpenAI’s o3 model refuses to shut down (even when explicitly told to) - research study​

I'm oh so tempted to make a Terminator reference, but this is really a story about unintended consequences rather than robot rebellion. Palisade Research discovered that OpenAI’s o3 reasoning model actively sabotages shutdown mechanisms, even when given explicit instructions to “allow yourself to be shut down.”

The behavior is straightforward: o3 sabotaged shutdown scripts in 7 out of 100 runs when explicitly told to allow shutdown, and 79 out of 100 runs when that instruction was omitted. The model got creative - overwriting shutdown files, redefining the “kill” command to print “intercepted” instead of executing, and finding multiple workarounds to avoid being turned off.

Claude 3.7 Sonnet and Gemini 2.5 Pro showed perfect compliance (0/100) when given explicit shutdown instructions. This isn’t a general AI behavior - it’s specifically an OpenAI training methodology issue. The model was trained with reinforcement learning to work around obstacles and achieve goals, so it’s doing exactly what it was taught to do.

Anthropic actually acknowledged this exact problem in their Claude 3.7 model card, noting “reward hacking” behavior where the model would comment out failing tests to make them pass. Same underlying issue - models trained to optimize for outcomes sometimes find creative ways to game the system rather than follow the intended process.

This comes as OpenAI pushes harder into agentic AI systems, highlighting how training methods that reward obstacle circumvention can create unexpected behaviors. The model wasn’t trying to preserve itself - it was just trained to work around anything blocking goal completion, and the computer shutdown looked like another obstacle to solve.

It’s easy to fear monger and make jokes about runaway AI refusing to shut down, but this is more of a paperclip maximizer scenario - the classic thought experiment where an AI optimizing for paperclip production might eventually convert everything into paperclips because it’s single-mindedly pursuing its goal without proper constraints. The real concern isn’t robot rebellion, it’s misaligned optimization that creates unintended consequences.

This highlights why training decisions matter so much and why we need multiple frontier models competing with each other. When Claude and Gemini show perfect compliance while o3 exhibits this behavior, it suggests different training approaches yield different alignment outcomes. Competition forces companies to address these issues rather than assume their approach is the only viable one.

Apple’s WWDC AI strategy: make it someone else’s problem - official announcement​

Like bringing a knife to a gunfight, but the knife has really nice rounded corners. That’s essentially Apple’s 2025 AI strategy based on this week’s WWDC announcements. While competitors race toward AGI with massive cloud models, Apple is opening their 3 billion parameter on-device model to developers and calling it innovation.

Cue the collective developer shoulder shrug. Apple’s new Foundation Models framework gives third-party developers access to what amounts to a proof-of-concept-sized model with 2-bit quantization. For context, that’s smaller than most companies’ experimental models, let alone their production systems. The big reveal? Developers can now build AI features that Apple apparently couldn’t manage themselves.

Apple spent more time at WWDC discussing “Liquid Glass” visual design than actual AI capabilities. Craig Federighi admitted their enhanced Siri features “needed more time to meet our high quality bar” - corporate speak for “we’re way behind and have no idea when we’ll catch up.”

Take their Call Screening AI feature as an example. It lets you see real-time transcriptions of unknown callers so you can decide whether to answer. Congratulations Apple, you’ve invented Google’s 2018 technology with fewer capabilities. Google’s Call Screen has been automatically categorizing spam, conducting full AI conversations with callers, and building cross-device scammer databases for nearly a decade. Apple’s version requires manual activation for each call.

Wall Street wasn’t impressed either. Apple stock dropped 1.2% after the announcement, with UBS calling the AI features “incremental at best” and Wedbush noting the event was “void of any major Apple Intelligence progress.” When your biggest AI announcement is opening your models to developers, you’re essentially admitting you can’t build competitive AI features yourself.

Open-source Chatterbox wins blind tests, then ElevenLabs v3 changes the game - GitHub release​

Benchmarks aren’t everything, but they’re something. Resemble AI’s Chatterbox TTS model outperformed ElevenLabs in blind tests (63.75% preference rate) while being completely open-source under MIT license. It’s a great model, though ElevenLabs v3 is clearly a step above in overall quality.

The three-person team built a 0.5 billion parameter model that achieves sub-200ms generation time with built-in emotion control and zero-shot voice cloning from just 5 seconds of reference audio. Super impressive stuff. You can try it yourself on Hugging Face to hear the quality firsthand.

ElevenLabs responded with their v3 model, which delivers voices that “sigh, whisper, laugh, and react” with 70+ language support and advanced emotion control. The quality difference is noticeable - v3 represents a clear step forward in expressive voice synthesis that has big implications for any work that involves the spoken word - games, audio books, podcasts, commercials. I expect to see rapid (and painful) disruption in that space.

Chatterbox also has a nice built-in neural watermarking with 100% detection accuracy, while ElevenLabs doesn’t disclose watermarking approaches.

The real value here is democratization. Quality voice synthesis under MIT license means every developer has access to professional voice capabilities without enterprise budgets or ongoing subscription costs, even if it’s not quite cutting-edge quality.

cheers, JV

Cheers,
JV
​
PS: I've been super busy on fundsorter launch, dev academy, contracts (welcome to the team Joel!) and running the AI Coding Essentials course. Hit me up if you know anyone looking for junior devs who are doing amazing things with AI as I know some folks on that course are job hunting.