o3 is a Filthy Liar, Google's spree, & More

TLDR;

This week, we first take a look at Google’s ongoing model streak and a disheartening reality about current AI assistants, the sycophancy disease. We then move on to discuss OpenAI’s potential acquisition of Windsurf, and discuss AI as a tool for job cutting.

Finally, we look at agentic frameworks and, in what is the biggest news of the week for me, o3’s deeply concerning deceiving nature.

MODEL RELEASES

Google Continues to Deliver

Google has officially entered the ‘hybrid model’ arena, models that behave as both non-reasoning and reasoning models depending on the user's request, with Google 2.5 Flash.

This model, a smaller version of Google’s leading Gemini 2.5 Pro, is extremely cost-effective to run, with an unbeatable price tag of $0.15 per million input tokens and $0.60 or $3.50 per output token (depending on whether you activate thinking or not).

Importantly, unlike our OpenAI friends, these guys openly compare their models with other labs, and the short summary is that it’s a model fairly in the o4-mini arena while being considerably cheaper.

TheWhiteBox’s takeaway:

As we discussed yesterday, this model further consolidates Google’s dominance of the performance-to-cost Pareto frontier (image thumbnail).

As for how their hybrid mechanism works, as it can be toggled, they are probably using different special tokens to elicit different thinking behavior. In layman’s terms, the underlying orchestrator appends the token <think> as a signal for the model to generate chains of thought. Sounds like magic but it’s literally training the model on several sequences of the style ‘<user input><think>chain of thought</think><answer>’.

Interestingly, it seems some labs may be eliciting this hybrid behavior at the context level. For example, you can define Claude’s behavior, at least in their Claude Code product, by using different prompt signals like telling it to ‘ultrathink’, ‘megathink’, or ‘think harder’.

Funnily enough, the idea that hybridism is elicited via simple word appends suggests that models can only know the level of thinking a problem requires based on external signals (via a system prompt the product team wrote in your behalf or simply you telling it).

To me, this kind of defeats the idea that models are developing “meta-awareness” that some proclaim they have; it’s not like the model is identifying the need to think more or less, and it still requires external support in some form. I’m not suggesting such meta-awareness won’t be possible, but it’s funny how we are consistently gaslighted and the reality is much simpler.

MODEL ALIGNMENT

Claude the Merciful

Talking about Claude, in its desire to meet user’s needs, Claude deliberately plays worse in a tic-tac-toe game with humans if you tell it you’ve had a rough day. As seen above, the win rate of Claude models collapses.

Weird, right?

TheWhiteBox’s takeaway:

The chart above serves perfectly to illustrate two things:

Model labs still struggle balancing utility (being the best model possible) with behaving like a helpful assistant.

These models can’t be trusted and they are too dependent on user input.

And I’m not sure which one…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.