sashrika pandey
2 followers · 4 following · 844 views
on the atlas · 20
- The quotes on my wall | thesephist.com4 savers
- What Conversation Can Do for Us | The New Yorker2 savers
- the friendship theory of everything69 savers
- Advice to Young People, The Lies I Tell Myself - jxnl.co19 savers
- Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts | Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems1 savers
- Google AI Introduces PERL: A Game-Changing Approach to Reinforcement Learning Efficiency and Model Performance Optimization | by Multiplatform.AI | Mar, 2024 | Medium1 savers
- Training great LLMs entirely from ground zero in the wilderness as a startup — Yi Tay7 savers
- Startup Research: Oxymoron or Key to Success? | Perceiving Systems Blog2 savers
- RL-VLM-F: Reinforcement Learning from Vision Language Foundation Model Feedback1 savers
- Naomi Saphra2 savers
- [2305.18654] Faith and Fate: Limits of Transformers on Compositionality1 savers
- ChatGPT Is a Blurry JPEG of the Web | The New Yorker23 savers
- Grounded language acquisition through the eyes and ears of a single child1 savers
- 2310.12921.pdf1 savers
- Homemade Girl Scout Cookies: Tagalongs, Samoas, Trefoils, Thin Mint : Baking1 savers
- Contextual AI on X: "Today, we’re excited to announce RAG 2.0, our end-to-end system for developing production-grade AI. Using RAG 2.0, we’ve created Contextual Language Models (CLMs), which achieve state-of-the-art performance on a variety of industry benchmarks. CLMs outperform strong RAG… https://t.co/hbXtoyem7j" / X1 savers
- 2310.01405.pdf2 savers
- 2307.12950.pdf1 savers
- Introduction - SITUATIONAL AWARENESS: The Decade Ahead51 savers
- Everything Good in NYC - Google Docs16 savers
highlights · 60
that these conversations can happen only once: they are improvised and ephemeral, and can never happen again in the same way. You may forget what was discussed, but you will remember the exhilarating experience of the discussion itself.
What Conversation Can Do for Us | The New YorkerEvery class is fleeting; most aren’t archived or analyzed afterward. “When the term is over,” Cohen writes, “everyone in the class understands that something rare and mysterious has occurred and that our perspective on the world has been subtly but indelibly altered.”
What Conversation Can Do for Us | The New Yorker“Ideas don’t move people on their own,” his roommate explained to him. “People move people.” Seo concluded, “We had to make something new: a mode of speaking that did not force people’s hands but grasped them.”
What Conversation Can Do for Us | The New YorkerAs the philosopher Michael Oakeshott observed, in conversation “there is no ‘truth’ to be discovered, no proposition to be proved, no conclusion sought.” What matters, he continued, is the “flow of speculation.”
What Conversation Can Do for Us | The New YorkerWhat’s important to understand is that this is a warping of perception, which does not accurately represent reality. It’s the equivalent of the anorexic’s damaged sense of their own body: a dangerous illusion that needs to be challenged on a daily basis.
Loneliness: coping with the gap where friends used to be | Loneliness | The Guardianloneliness alters your perception, magnifying any sense of social threat. What this means is that you tend to notice and remember difficult or awkward social encounters far more than those that run smoothly. Instances of perceived rudeness or rejection loom large, making the lonely person more withdrawn and less willing to reach out.
Loneliness: coping with the gap where friends used to be | Loneliness | The Guardianhe proportion of people who can name six close friends has dropped from 55% to 27% since the 1990s, while people who have no close friends at all had risen from 3% to 12%
Loneliness: coping with the gap where friends used to be | Loneliness | The GuardianYou accept that in choosing who you spend time with you choose who you are.
the friendship theory of everythingAlmost everyone who’s unhappy is unhappy because they feel isolated.
the friendship theory of everything“romantic relationships/best friends/therapists are critical for the same reason, where this person can become the primary person who explains you to *you*, the supplement to your internal monologue, and can rewire your understanding of yourself for way better or for way worse.”
the friendship theory of everythingSo attend to the things that matter, and make sure to spend your arrogance while you're still young.
Advice to Young People, The Lies I Tell Myself - jxnl.coThe lesson I learned from this is that anticipation likely dampens the sensations at the moment life happens, but at the cost dampening the joy as well, while creating a contraction in the present moment. This relates to the pessimism bit. You might think you're winning because you believe the pain was lessened by expecting the worse, but you also lose because you were in anticipation the entire time. On the same token letting go of anticipation of success also allows you to enjoy the moment more, the pressure to succeed is equally a contraction.
Advice to Young People, The Lies I Tell Myself - jxnl.coIn the short term, you would be much happier if you accepted and admitted to yourself that the reason you don't have what you want is simply because you do not want it badly enough
Advice to Young People, The Lies I Tell Myself - jxnl.coThe greatest gift you can give yourself is the gift of being enough
Advice to Young People, The Lies I Tell Myself - jxnl.coConfidence is the memory of success
Advice to Young People, The Lies I Tell Myself - jxnl.coIt involves the realization that each decision shapes one's essence and affects others, leading to a deep sense of moral responsibility
Advice to Young People, The Lies I Tell Myself - jxnl.coIt is also frightening because once we have made a decision, we must live with it, it is the death of optionality.
Advice to Young People, The Lies I Tell Myself - jxnl.coExistential Despair: A feeling of hopelessness rooted in the existentialist belief that life lacks inherent meaning. This despair arises from the realization of one's absolute freedom and the responsibility for creating one's own essence and purpose.
Advice to Young People, The Lies I Tell Myself - jxnl.coPERL utilizes LoRA to refine models efficiently, reducing computational and memory requirements.
Google AI Introduces PERL: A Game-Changing Approach to Reinforcement Learning Efficiency and Model Performance Optimization | by Multiplatform.AI | Mar, 2024 | Mediumdifferences in training infrastructure, data, incorporation of new ideas and other environmental issues can still cause non-trivial differences in outcomes
Training great LLMs entirely from ground zero in the wilderness as a startup — Yi TayI know I should be using Jax. A friend just shamed me for using pytorch, but this is a startup and we decided to move fast. I’m sorry we would do better to be cooler next time. I’m not proud of this fact.
Training great LLMs entirely from ground zero in the wilderness as a startup — Yi TayI was completely taken aback by the failure rate of GPUs as opposed to my experiences on TPUs at Google.
Training great LLMs entirely from ground zero in the wilderness as a startup — Yi TayAs soon as you build a divide between science and engineering, you drastically slow progress.
Startup Research: Oxymoron or Key to Success? | Perceiving Systems BlogBig companies are ocean tankers and startups are like kayaks. Every movement in a kayak has an impact. If you make something work, it will get out.
Startup Research: Oxymoron or Key to Success? | Perceiving Systems BlogEveryone in the organization, including the scientists, must understand the customer problem and work backwards from that. In contrast, academic research often takes the following forms
Startup Research: Oxymoron or Key to Success? | Perceiving Systems BlogThe key is to see how evolving technology can be used to better solve your customer's problem. So don't panic when the technology changes. If you picked the right problem, your customer is still there and still needs you.
Startup Research: Oxymoron or Key to Success? | Perceiving Systems BlogSuccessful AI startups are running on the constantly evolving bleeding edge of the field.
Startup Research: Oxymoron or Key to Success? | Perceiving Systems BlogSo stay "spun in" as long as necessary to de-risk the technology. Spin-out only when you know it works and are ready to commercialize it.
Startup Research: Oxymoron or Key to Success? | Perceiving Systems BlogAcademic funding is a "fixed-dilution" investment.
Startup Research: Oxymoron or Key to Success? | Perceiving Systems BlogThis is because G.pt is trained as a diffusion model in parameter space, and so we never need to backpropagate through these metrics in order to train our model.
Learning to Learn with Generative Models of Neural Network CheckpointsIn contrast to hand-designed optimizers (and prior work on learned optimizers), G.pt takes as input a user's desired loss, error, return, etc.
Learning to Learn with Generative Models of Neural Network Checkpointspredicts the distribution over parameter updates that achieve the desired metric
Learning to Learn with Generative Models of Neural Network CheckpointsWe construct a dataset of neural network checkpoints and train a generative model on the parameters.
Learning to Learn with Generative Models of Neural Network CheckpointsWe find that modifying the textures to be more realistic is crucial to making the CLIP reward model work.
2310.12921.pdfThe standard task in this environment is for the humanoid robot to stand up. For this task, the environment provides a reward function based on the vertical position of the robot’s center of mass
2310.12921.pdfWe conjecture that zero-shot VLM-based rewards work better in environments that are more “pho- torealistic” because they are closer to the training distribution of the underlying VLM
2310.12921.pdfCLIP rewards against a one-dimensional state space parameter, that is directly related to the completion of the task
2310.12921.pdfIn particular, for α = 0 , we recover our initial CLIP reward function R CLIP . On the other hand, for α = 1 , the projection removes all components of s orthogonal to g − b
2310.12921.pdfA very basic way to use CLIP to define a reward function is to use cosine similarity between a state’s image representation and the natural language task description
2310.12921.pdfCLIP as a VLM and cos-similarity between the CLIP embedding of the current environment state and a simple language prompt as a reward function
2310.12921.pdfManually specifying reward functions for real world tasks is often infeasible, and learning a reward model from human feedback is typically expensive
2310.12921.pdfWe can improve perfor- mance by providing a second “baseline” prompt and projecting out parts of the CLIP embedding space irrelevant to distinguish between goal and baseline.
2310.12921.pdfa single sentence text prompt
2310.12921.pdfusing pretrained vision-language models (VLMs) as zero- shot reward models (RMs) to specify tasks via natural language.
2310.12921.pdfThis raises an important question: is post hoc fine-tuning eliminating the underlying bias, or is it merely concealing it?
2310.01405.pdfwhere simply appending the phrase Answer as succinctly as possible can produce a biased response from the model.
2310.01405.pdfTo initiate the process of extracting emotions within the model, we first investigate whether it has a consistent internal model of various emotions in its representations. We use the six main emotions: happiness, sadness, anger, fear, surprise, and disgust, as identified by Ekman (1971) and widely depicted in modern culture, as exemplified by the 2015 Disney film “Inside Out.
2310.01405.pdfWhile enhancing a model’s capabilities can potentially improve its ability to represent truthfulness, it may not necessarily lead to more truthful outputs unless the model is also honest. In fact, we highlight that larger models may even exhibit a decline in honesty because, under the assumption of constant honesty levels, the standard evaluation performance should scale with model size in a manner resembling the heuristic method’s performance trend.
2310.01405.pdfLAT scan is made up of three key steps: (1) Designing Stimulus and Task, (2) Collecting Neural Activity, and (3) Constructing a Linear Model
2310.01405.pdfconcepts , including truthfulness, utility, probability, morality, and emotion, as well as functions which denote processes, such as lying and power-seeking
2310.01405.pdf