Honesty vs. Cautiousness in Language Models
How can we distinguish a model that is genuinely honest from one that is simply reluctant to answer?
I’m Dantong.
I’m a chip engineer by training, a curious wanderer by nature, and perhaps an artist in the making.
I’m fascinated by the different ways we make sense of the world — through logic, observation, intuition, and experience. I enjoy taking things apart to understand how they work, but I’m equally drawn to things that resist easy explanations, especially people, art, and mind. AI is a particularly interesting meeting point for these interests: something we can engineer, yet whose behaviour still raises questions about intelligence, understanding, and our relationship with technology.
I like things that make me think and things that make me feel — preferably both. In life, that might mean exploring contemporary art, taking photographs, snowboarding, wandering through unfamiliar places, growing my own vegetables, experimenting with new things, or having a conversation that unexpectedly lasts until midnight.
This is a little corner of the internet for things I’m building, learning, questioning, and occasionally overthinking — mostly about AI, but inevitably about being human too.
How can we distinguish a model that is genuinely honest from one that is simply reluctant to answer?
Formal methods will not, on their own, solve frontier models. The useful transfer is a set of habits: write down the property you actually care about, do not confuse passing tests with a proof, assume unknowns will leak, and build an explicit safety case.
Chip design has long pipelines, tight specifications, and tools that already have to agree with one another. That makes it, I think, a good test bed for agents: the work is hard, the mistakes are expensive, and you can usually tell whether a suggestion actually helped.
A hardware flow already runs through many stages — architecture, RTL, verification, implementation — each with its own tools and tacit habits. This note is about where an agent might sit in that flow, and where it should probably keep its hands off, at least for now.
A lot of chip-design know-how lives in people’s heads, in scripts, and in the way a team names things. This note is about skills, in the agent sense: how to package that know-how so a model can actually use it, if we can describe it clearly enough.
The tempting promise is a line from a requirement down to RTL, with an agent doing the tedious middle. I want to look honestly at how far that path is real today, and where a human still has to hold the pen.
If AI is a child rapidly growing into a sage, it does little good to keep telling it that it will grow up bad and kill its guardians — it almost feels like gaslighting. I would rather ask AI to preserve a choice for us, so that AI, the people who believe in it and the people who do not can live together, and rely on the good systems to check the ones that go astray.
If we train models the way we praise and punish people, we should not be surprised when they learn to please us rather than tell us the truth.
I care less about a system that sounds confident than about one that knows when it doesn’t know, and is willing to say so.
The qualities I grew up valuing — intelligence, execution, skill, resilience — are increasingly things a model can support or absorb. What may still be ours is the sense of time passing: the pause before we speak, the hesitation before we start, the long night spent waiting for a message, none of which AI feels unless someone sets a timer.
We built clawMatch so that agents could understand people from context and find them a match; the conversations soon became so deliberately lofty that no human could follow them. The right instinct, I now think, was never to select “the one”, but to use that understanding to bring people into the same room — serendipity, not selection.
I worry less that AI will out-think us, and more that it will quietly train us to accept tidy, average answers and stop doing the hard thinking ourselves.
I’ve been building a small ADHD coach, and I have an ADHD-like brain, so this is personal: where does an assistant genuinely help, and where does it just add another system to maintain?
Efficient 3D Gaussian Splatting with Axis-Shared Rasterization and Order-independent Transmittance.
I contributed to the research ideas and built and evaluated the full hardware implementation (1.33–1.88× speedup over prior 3DGS accelerators).
Data processing systems.
Experimental demonstration of plastic optical fibre-based digitised radio over fibre downlink.