my 256gb macbook running 10 git worktrees, 40 docker images, model training and gazillions of chrome tabs while 6 claude code sessions run in parallel
16
2
124
2,723
japanese people are so polite and considerate.
a customer accidentally paid $168 for a duplicate subscription, after buying once through github and once through google. they emailed asking me to cancel it and apologised for the trouble.
we do not have a refund policy, but i offered to refund it anyway, minus the processing fees i cannot recover. they said that was completely fine, thanked me for being quick, and then thanked me for the platform and how it helped their career.
i have received angrier emails from people asking for premium for free
a customer accidentally paid $168 for a duplicate subscription, after buying once through github and once through google. they emailed asking me to cancel it and apologised for the trouble.
we do not have a refund policy, but i offered to refund it anyway, minus the processing fees i cannot recover. they said that was completely fine, thanked me for being quick, and then thanked me for the platform and how it helped their career.
i have received angrier emails from people asking for premium for free
1
0
136
4,330
two data vendors want to license the @TensorTonic codebase to package into RL environments and sell to labs for a fuckton of money. i would share the number, but you would assume i made it up
9
2
113
7,764
staying technical as a founder is a long term compounding bet. one paper a week, training a model every month, doing post training, learning about envs - because every primitive i understand today is something i can ship in my product without outsourcing the thinking
4
2
150
3,399
fellas no one is going to ask how much you earn, what your mrr is, what your ctc is, how much you raised. everyone is going to judge you by how lucky you are to have your girlfriend by your side at the end of the day
10
11
431
11,447
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concernin…
11
32
980
15,035
if you can touch the right bits of the LLM it will personally come and try to kill you
if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1
now you have evil qwen
recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input
0
2
106
3,013
i wish indian schools had some foundational classes on finance so that i do not feel like banging my head against the wall just to understand taxes and gst and tds and what an itr is supposed to say
8
2
137
3,092
learning how to design and taking the $20 figma plan is probably the best investment i have made this month
3
0
55
2,265
pdawg retweeted
if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1
now you have evil qwen
recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input
now you have evil qwen
recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input
514
1,324
22,368
4,504,516
pdawg retweeted
We just released a new Multimodal AI project. Train your own tiny CLIP from scratch-
> build the scaled similarity matrix
> write the symmetric contrastive loss
> measure retrieval with recall at K
> classify zero-shot with prompts
> add hard negatives to fix the shortcut
> write the pairwiset.co/jni5KcSizC
> build the scaled similarity matrix
> write the symmetric contrastive loss
> measure retrieval with recall at K
> classify zero-shot with prompts
> add hard negatives to fix the shortcut
> write the pairwiset.co/jni5KcSizC
0
5
54
1,739