Pinned Tweet
Data Ethics course: ethics.fast.ai
Deep Learning course: course.fast.ai
Data Science blog: fast.ai/topics/
Diversity blog: medium.com/@racheltho
NLP: bit.ly/fastai-nlp
Talks: bit.ly/rachel-videos
Medicine, AI, & Bias: bit.ly/MLmed
Deep Learning course: course.fast.ai
Data Science blog: fast.ai/topics/
Diversity blog: medium.com/@racheltho
NLP: bit.ly/fastai-nlp
Talks: bit.ly/rachel-videos
Medicine, AI, & Bias: bit.ly/MLmed
34
901
3,112
"The designer who tries to eliminate the [human] operator still leaves the operator to do the tasks which the designer cannot think how to automate. This causes problems, as they are left w/ an arbitrary collection of tasks & little thought was given to providing support" 1/
2
13
62
3,035
"Automation of industrial processes may expand rather than eliminate problems with the human operator.
"Even highly automated systems need human beings for supervision, adjustment, maintenance, expansion & improvement." 4/
"Even highly automated systems need human beings for supervision, adjustment, maintenance, expansion & improvement." 4/
1
0
4
610
"It is impossible for even a highly motivated human being to maintain effective attention towards a source of information on which very little happens."
"This means that it is humanly impossible to carry out the basic function of monitoring for unlikely abnormalities." 5/
"This means that it is humanly impossible to carry out the basic function of monitoring for unlikely abnormalities." 5/
0
0
10
587
Rachel Thomas retweeted
Specifics: Saying "don't do" something doesn't work as well as you'd hope because LLMs aren't as good at processing negation as you'd hope (arxiv.org/abs/2209.12711 , arxiv.org/abs/2601.08070). We've seen this cause code to be wiped out multiple times (fortune.com/2025/07/23/ai-…, theguardian.com/technology/202…) 2/
3
14
77
4,880
"How we perceive the world and how we act in it are products of how and what we remember." (Foer)
"Intuition is built either by constantly working on a single topic for very long or by actively learning (memorizing) new concepts. It seems that LLMs discourage both." 2/
"Intuition is built either by constantly working on a single topic for very long or by actively learning (memorizing) new concepts. It seems that LLMs discourage both." 2/
0
4
17
989
"[Prior to LLMs], the concepts you see every day find their way into your long-term memory, if only by rote repetition... [With LLMs] we keep showing new concepts to the brain constantly so it never has the chance to commit them to long-term memory." 1/
Just finished reading "Remember, Remember" and "Moonwalking with Einstein" by Ed Cooke and Josh Foer respectively.
The are books about memory. The first one shows you the magic, the second one how it's done.
They made me realize that a lot of the problems that LLMs bring to the table, like reduced problem-solving skills, might actually be problems derived from memory impairment.
Read more: medium.com/@pol.avec/revi…
6
1
33
3,043
"We argue for a reorientation from autonomous to human-centered coding agents: systems designed not only to complete tasks, but to collaborate effectively with people."
@ZhiruoW @jyangballin @KLieret
arxiv.org/abs/2608.12355 1/
@ZhiruoW @jyangballin @KLieret
arxiv.org/abs/2608.12355 1/
10
23
103
7,250
"The goal of developing coding agents is not to replace human labor, but to augment complementary human strengths such as initiative & judgment.
"Human-centered interaction is not a temporary workaround for immature models: it enables new forms of collaboration." 3/
"Human-centered interaction is not a temporary workaround for immature models: it enables new forms of collaboration." 3/
1
0
7
861
"If we continue optimizing solely for autonomous coding agents, we will produce just that. Better collaborators will not emerge for free.
"The research community must decide whether to optimize for leaderboards or the people who actually use these systems as well." 4/
"The research community must decide whether to optimize for leaderboards or the people who actually use these systems as well." 4/
0
0
5
641
A few papers & benchmarks on LLM failure to follow instructions 🧵
I. Can LLMs Unlearn Stubborn Training Conventions?
“LLMs often exhibit cognitive inertia, struggling when tasked with instructions that explicitly deviate from their training norms” arxiv.org/abs/2509.04292 1/
I. Can LLMs Unlearn Stubborn Training Conventions?
“LLMs often exhibit cognitive inertia, struggling when tasked with instructions that explicitly deviate from their training norms” arxiv.org/abs/2509.04292 1/
5
12
48
5,266
"In zero-shot settings, preference following accuracy falls below 10% at merely 10 turns (~3k tokens) across most evaluated models.
"Even with advanced prompting and retrieval methods, preference following still deteriorates in long-context conversations." 8/
"Even with advanced prompting and retrieval methods, preference following still deteriorates in long-context conversations." 8/
1
0
1
617