patterning personal intelligence
residual stream
- Tweets 2,081
- Following 1,579
- Followers 238
- Likes 9,329
Pinned Tweet
made a mac app for myself and it got good enough to share
it's called arden. my personal assistant, if you will: it remembers and does stuff for me, and tells me what's happening and what i need to do. and do most of it.
spent most of this year on it and i'm sharing it just because i can lol
p.s. thanks claude for this (and other) video haha
3
10
5,556
where does "blue" live?
apparently, in qwen3-4b there is no single blue neuron
each blue thing takes its own path:
- the samsung logo goes through a "korea" neuron
- the twitter bird through a "tweets" neuron (haven't checked X tho haha)
- police uniforms through an "NYPD" neuron
and all these paths end in a small late stage that says the word "blue"
---
by "korea" neuron i mean this: a neuron that fires on "Samsung" and makes "Korea" and "Seoul" more likely
1
63
tim ganiev retweeted
Can interpretability shape training?
@TIME covers our latest work on mechanistic shaping and why our team is optimistic. Read the story online or in TIME's upcoming print edition.
time.com/article/2026/11/05/…
5
17
110
4,991
tim ganiev retweeted
We’ve been hard at work trying to make models work with harnesses better. The best thing @yb2698 found: “harness adapters”. Small LoRAs trained with RL inside the actual agent harness they’ll be used with.
Same base model. Different adapter for pi, opencode, hermes, etc. Why?
Simply because the same model behaves very differently depending on the tools, prompts and control flow around it. Instead of asking one set of weights to be optimal everywhere, we can specialize a tiny part of the model for each environment.
Early results are pretty encouraging: on held-out tasks, the adapters make fewer tool calls than the base model while keeping or improving accuracy.
And because they’re LoRAs, the operational side is surprisingly cheap too.
For one LFM2.5-24B-A2B experiment, we evaluated 9 checkpoints / 1,350 agent sessions on a single H200 in under an hour. The total: ~$5!
Merging and serving every checkpoint separately would have cost around $90.
There’s another important piece: RL agents will eventually try to hack the reward.
So every tool call goes through `safeagents`, a deterministic rule-based guard that blocks things like editing tests, tampering with rewards or bypassing the sandbox before they execute.
And the whole loop can start from real usage:
your coding sessions -> verifiable environments -> RL -> a better adapter for your agent.
The model doesn’t have to change completely.
Sometimes it just needs to get better at the harness it actually lives in.
huggingface.co/blog/harness-…
37
34
169
26,090
my fav benchmark
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/ink…
2
315
finally someone using mechinterp properly
We cut through the mystery around interpretability in training.
A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities.
We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment.
But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵
1
16
1,496
tim ganiev retweeted
If your cocaine dealer shorted you and didn’t supply you with the full amount of product, contact us at 336-373-2222 to file an official complaint.
5,593
34,258
357,931
18,113,224
tim ganiev retweeted
people are like “why do LLMs always think they’re in an eval lol” and then have bad dreams about forgetting to turn in a homework assignment 10 years ago
78
352
6,424
134,133
just caught myself describing concepts to my agent in pseudo-code because it's simpler and denser than english
full circle
1
2
1,718
tim ganiev retweeted
You should apply to this :)
Fun personal fact: I applied to OpenAI residency like 8 times.
Applications are now open for the Anthropic Residency, a paid, fixed-term research program for academics, technical founders, and staff-level engineers who want to make AI safe and beneficial for everyone. No ML background is required.
Apply by November 22: alignment.anthropic.com/2026…
51
25
1,483
191,427
wow youtube.com/watch?v=L1OBRtfP… from @leerob have so many cool details in there
i like the part about context pre-warming, it's the part you usually don't need in other agents but makes interaction with agent really-really smooth
2
1,468
tim ganiev retweeted
we gave qwen emergent misalignment by training it on bad advice
in two separately trained models, the same neurons lit up on its misaligned answers
we looked them up in my neuron atlas, and one theme kept coming up: power 🧵
13
20
209
10,252
> fine-tuning is cope for a slow training. do not add post-training and delete all existing loras. instead, redesign the full training pipeline to be faster
some things I often say to claude when doing perf optimization work:
> how do we rewrite this to better leverage data-oriented design principles
> can you replace those hash tables with a flat array? can we make the pointers an index into the array? can we pack the most commonly used data into upper bits of the index?
> caching is cope for a slow design. do not add caches and delete all existing caches. instead, redesign it to be faster
> should this be a structure of arrays instead of an array of structures?
2
1,297
