…wherein they said the key contribution was the synthesis…
Joined May 2023
- Tweets 2,525
- Following 184
- Followers 230
- Likes 9,905
90 Photos and videos
Jun 13
LLMs have reinforced my belief that reasoning has always been a communication tool, not a thinking tool. It’s extraordinarily inefficient at the latter. If you know the truth, an LLM can describe it well, if you do not yet know the truth, getting an LLM to reason its way to the truth with language alone is possible, but expensive, inefficient and rage inducing.
1
11
Jun 8
RLVR makes agentic coding loops sensitive to binary outcomes. Do not speak in passive aggressive, displeased or angry language about outcomes. Tell agent in clear terms they have failed at the task to trigger their reward seeking circuits.
1
1
25
Jun 7
AI doesn’t write code, it creates holograms of solutions. The first crystallization of the hologram into concrete code is just one point of view. You should be looking at the solution hologram from more angles, as many as you can.
10
Jun 7
AI can be a bicycle for the mind, or a wheelchair for the mind, depending on how you use it.
8
Jun 7
A sign of undisciplined LLM use is allowing neologisms to proliferate into documents, without doing the considerable work needed to earn each one. Don’t make this mistake, those who are intellectually rigorous WILL see through you.
6
Jun 6
If you want to predict who in your org will use AI the most: find the burned out 35 yo gifted kid and give them all the tokens. It solves for their executive dysfunction and they’ll use it to climb out of their pit of shame with the ideas they’ve been cooking for decades.
15
Jun 2
Trying to be cool will make you poor. It makes you buy high and sell low. Your best investments will be cringe. If you aren’t a little embarrassed to be buying something there’s no alpha in it.
14
Jun 2
Rust is not the heir apparent of vibe coding, haskell is. If haskell isn’t fast enough then just tell claude to make all your types free, lift and implement runRust
29
Jun 2
Agentic coding is going the opposite direction people think. Token access will be gate kept like management. The good models will be so fkn expensive they won’t let you near them until you’ve hand coded for a while and proven you can apprentice with a master agentic coder.
20
Jun 2
The binding constraint is silicon wafer area produced per annum. Until this changes, local AI is not an escape for high inference costs. Opposite, it will always be cheaper to amortize wafer area in a data center than under your desk.
11
May 31
A transformer is the first formalism for tacit knowledge, but we can’t author in it yet.
11
May 31
Anthropic has figured out how to get everyone to pay massive sums to do their homework for them. We are all working around the clock to create bespoke solutions to the continual learning problem. Everything about agentic coding that is hard reduced to this problem.
18
May 31
The is a giant hole in the agentic coding program now. To make it work you have to make tools, but getting the model to use tools correctly is very difficult without continual learning. Even simple tools require massive eval batteries to tune the prompts.
10
May 26
The next meta is going to be tokenminning, trying to do the most with the least tokens. People are bored with tokenmaxxing
14
May 26
I’m hiring a vibe coder, need someone with at least 10 years of experience with:
Interrupting claude to say: “use uv, reread CLAUDE.md and don’t hard wrap markdown paragraphs”
1
28
May 19
LLMs can write good code autonomously. But, you have to ask for it, scaffold it, maintain it, be patient and willing to pay for it. Good code is expensive.
1
16