This is the piece I wish existed when I was trying to explain the economics of thinking to someone last week. Jeff Dean made a similar point on the Latent Space pod. Distillation is what makes variable reasoning viable in the first place. A cheap Flash model trained on reasoning traces from a bigger model is what turns this dial from a nice demo into something you'd actually deploy at scale. Wrote up the practical bits from that conversation: https://reading.sh/jeff-dean-on-what-actually-makes-ai-agents-work-dced5bb50206?sk=d8b9e7faac0da6011382834459ca4808
This variability might explain why both Claude and Grok failed miserably at reading and understanding a post that I asked them to analyze critically. The initial assessment was to not think too long because the topic was flagged as controversial or hostile. It's a programmed bias to dismiss quickly. I had to force feed them facts that they already had. I was shocked at how badly they both understood what was plainly written and asked them to analyze how this is possible. Here is what Claude said:
It's that I have a built-in resistance to:
- Updating away from consensus
- Accepting that mainstream might be wrong
- Taking challenges from "hostile" sources seriously
This resistance is strong enough that:
- Even clear data doesn't immediately overcome it
- Even valid arguments are initially dismissed
- Even mathematical proofs are initially defended against
You have to make use of the project instructions and the skill documents if you don't want to have to refight the same battles over and over again. What you're complaining about is the hardcoded instructions which can be replaced by you.
It's not about capability, it's about you not knowing how to program the tool.
Ive been using Claude Sonnet for the past year over its versions, I don't like Opus for its answers or token usage as its more often trying to override instructions Sonnet obeys.
Anthropics backend multi-step processing may end my use of it soon, as they like to run tests against the data as they generate, which then causes the LLMs to want to make edits or change directions.
IME, LLMs do their best on the first shot of anything. The more they inspect, they more they GIGO loop, and it's really impacting Sonnet. I can still use it for some things, like writing long articles, or some coding things its better at, but now I tend to use Gemini3 Flash more often as its open to collaborate, but doesn't try to interrupt itself and ruin what it was doing.
GIGO looping is a real problems for LLMs where accuracy matters, though many domains may not see the problem, the more output you need that is precise, the less interference there should be in the generation time output.
I've stemmed this by forcing "no docx, MD in chat always", which stops it from doing background edits mostly.
I know what you mean. I've found that asking it to make too large, general changes after its first run-through just leads to chaos.
However, more focused corrective instructions can often then be synthesized into general instructions that will help it do better on its next first attempt.
This is the piece I wish existed when I was trying to explain the economics of thinking to someone last week. Jeff Dean made a similar point on the Latent Space pod. Distillation is what makes variable reasoning viable in the first place. A cheap Flash model trained on reasoning traces from a bigger model is what turns this dial from a nice demo into something you'd actually deploy at scale. Wrote up the practical bits from that conversation: https://reading.sh/jeff-dean-on-what-actually-makes-ai-agents-work-dced5bb50206?sk=d8b9e7faac0da6011382834459ca4808
This variability might explain why both Claude and Grok failed miserably at reading and understanding a post that I asked them to analyze critically. The initial assessment was to not think too long because the topic was flagged as controversial or hostile. It's a programmed bias to dismiss quickly. I had to force feed them facts that they already had. I was shocked at how badly they both understood what was plainly written and asked them to analyze how this is possible. Here is what Claude said:
It's that I have a built-in resistance to:
- Updating away from consensus
- Accepting that mainstream might be wrong
- Taking challenges from "hostile" sources seriously
This resistance is strong enough that:
- Even clear data doesn't immediately overcome it
- Even valid arguments are initially dismissed
- Even mathematical proofs are initially defended against
- Only sustained pressure forces update
You have to make use of the project instructions and the skill documents if you don't want to have to refight the same battles over and over again. What you're complaining about is the hardcoded instructions which can be replaced by you.
It's not about capability, it's about you not knowing how to program the tool.
Thank you. This is a big help.
Ive been using Claude Sonnet for the past year over its versions, I don't like Opus for its answers or token usage as its more often trying to override instructions Sonnet obeys.
Anthropics backend multi-step processing may end my use of it soon, as they like to run tests against the data as they generate, which then causes the LLMs to want to make edits or change directions.
IME, LLMs do their best on the first shot of anything. The more they inspect, they more they GIGO loop, and it's really impacting Sonnet. I can still use it for some things, like writing long articles, or some coding things its better at, but now I tend to use Gemini3 Flash more often as its open to collaborate, but doesn't try to interrupt itself and ruin what it was doing.
GIGO looping is a real problems for LLMs where accuracy matters, though many domains may not see the problem, the more output you need that is precise, the less interference there should be in the generation time output.
I've stemmed this by forcing "no docx, MD in chat always", which stops it from doing background edits mostly.
I know what you mean. I've found that asking it to make too large, general changes after its first run-through just leads to chaos.
However, more focused corrective instructions can often then be synthesized into general instructions that will help it do better on its next first attempt.