February 2024

This week, I learned: Architecture.md is an emerging standard Managing wealth requires training. htmz is a fantastic way to load HTML into elements! Suguna Poultry is Using robots to walk in their farms, use sound and bird eyes and movement to predict birth health over 1-2 weeks Light on the back of the bird’s back AND face => lays eggs in 14 days, else takes days later (girls and mobile phones?) Teknoturf is using Gen AI to Improve prompts when teaching prompt engineering. Pronounce languages better, identifying which words Tamilians and Malayalis will mis-pronounce. Explore IRBlaster. It can control AC and can automatically increase temperature at night. My view: LLMs are general purpose and more capable than SLMs. They’ll win, like CPUs won over special-purpose chips. GPUs will optimize for LLMs and as usage grows, cost will fall. Andrej Karpathy’s summary of sharp edges in tokenization uses tiktokenizer to explain: Why LLMs can’t be used for spelling Why LLMs are better at English than other languages Why LLMs are bad at math Why SolidGoldMagiKarp is a single token Why trailing spaces are bad Why YAML tokenizes more efficiently than JSON ssyoutube.com: Just add “ss” to “youtube.com” on the video and you can download YouTube videos Discussions with Sachin, AMAT Microsoft said Indigo, Air India uses LLM based bookings Meta invested $70bn in GPUs. Sam Altman is investing $7tn! NVIDIA has a price PREMIUM not discount for bulk GPUs! AMD is the next company to watch for Numenta - Subutai Ahmad - deploys AI models on CPUs #TODO Read A Thousand Brains by Subutai Ahmad Sanjeev Sharma Swaayatt Robots: Autonomous driving in India Deepeigen: Education Rohan Shravan, Bangalore. Likes sharing knowledge. Amazing teacher. IIT KGP 2008. Interested in exploring quantum computing Tresa Motors, Inkers App, The School of AI AMAT is working on photon-based computing. science research models. AI for science. Like Google: Deepmind Genome, Microsoft: Metagen quantum: AMAT is actively in into this. Nagapati Banda is driving this John Kelly is predicting a ChatGPT moment in quantum in a few years Adobe express has a forever free video to GIF converter Edge workspaces let me keep the same tabs open across laptops! Command line interface guidelines RAWGraphs has a custom charts API that is worth learning from Python fastcore has decorators like @typedispatch, Self, etc. All image-to-text models on HuggingFace wddbfs mounts SQLite as a file system. I had a bit of trouble, maybe with Python package versions. Google is using LLM powered bug identification HuggingFace Chat Assistants has open source system prompts!! OpenHermes training dataset is available. 1M prompts! Jio has made IPL free. They make money on data and ads. That’s Scale! Daniel Dennett outsources thinking to students. Reviewing his books. BUT: I don’t take feedback. When someone sends a pull requests, I ignore it.

This week, I learned: Fine tuning makes economic sense only if the input tokens SAVED is twice the output token size on each call. Docker container memory usage on WSL2 docker stats frolvlad/alpine-glibc:alpine-3.17: 540KB ubuntu: 1MB (python3: +5MB) nikolaik/python-nodejs:python3.10-nodejs18-bullseye: 1.4MB (python3: +5MB) python:3-alpine: 612KB (python3: +7.5MB) python:3: 500KB (python3: +11.2MB) continuumio/miniconda3: 7.6MB (+6.5MB) Discussion with Vinu Yamunan Databuck by FirstEigen. Autolysis plus monitoring Quality council has the data steward (maintainer of each dataset) coming together with the uses on a weekly basis to understand what quality problems to users are facing. Data owners jaundice at a lower frequency to get an understanding #TODO Automate rules for data quality in our projects and intranet Convert a config rule into business language. Explain SQL. These are good use cases for llm’s Graph DBs are powerful for flexible data structures, but query generation needs AI or expertise. Check the Neo4J language cypher Explore storing SAME data in relational DBs AND in graph DBs / document DBs for different use cases Dallas rocketry challenge. Build a rocket that can take an egg to 800 feet exactly and land without breaking it Discussion with Karthik A #TODO Ask IIT students to do internship tasks. Use advent of code is a qualifying criterion Tata motors unionized DB admins for longevity. No one can take their jobs. Hires people who LIKE their jobs Rust gives me typing. It’s very efficient. Pola.rs is interesting but Pandas as good enough. Explore alerts from CCTV feeds. Karthik sends email alerts with pictures for: “Is the machine on or off”? for productivity “Are people not wearing helmets?” for safety at Cummins #TODO Integrate with WhatsApp. Use LLMs with function calling for responses Use expiring links (to pictures or content). It increases engagement Check Deno licensing. Is there a commercial clause? #ANS No - it’s MIT license Centre or excellence for zero emission tech at IIT. Karthik is part of it Explore auth0. 7000 users are free toml is part of the Python 3.11 standard library! If copilot writes code we don’t understand we are screwed. Hence expertise matters Discussion with Vikas Kedia #TODO Plan an AMA The mind becomes lazy with financial success. Vikas is treating his podcast as a startup Hire a professional videographer for your content Financial RoI in financial markets is the highest. Programming is high too but FS is even better “Performative power” – when you’re forced to perform, you get better ideas Observable 2.0 is an open source static site generator for data Python dataclasses SORA is OpenAI’s video generation model, and is stunning! If Appa comes to Singapore even for a week, he will feel better and can boast to his friends. At over 90, it may be better to move Appa to where I am since many of his friends would be no more and shops, doctors, etc can be managed and getting an independent house nearby is not hard. There is an SEZ in Gujarat where Indians can invest like in Mauritius without forex restraint Shubha: Media sites are moving away from Vickrey auctions to first-price auctions for ads. That’s because they send the auction price forward to a search engine and the winning second-price value can lose even though the owner is willing to pay more. Second-price auctions don’t work unless ALL bidders are in the SAME auction. Ad networks are a hierarchy of auctions! Gemini 1.5 launched. Fly.io offers GPU hosting and auto stop when they have nothing to do. Embeddings in random forest are very effective at classification – much better than dot product. To deploy apps with OAuth + templating support in a small Docker container, use Caddy Deno has native TypeScript, browser APIs, and compiles to multiple OSs Ruff is a MUCH faster flake8 Two pass generation is a clever technique to get multiple SEQUENTIAL answers in a single API request. For example the schema {'code', 'optimized_code'} will generate code and then optimize it. Unions in function calling allows flexible multi-step prompts in a single API.

This week, I learned: Dockerfile can have FROM scratch and you can add specific binaries rather than an entire OS. via Fine-tuning session by Dan. Notebook Example of fine-tuning Mistral. Consumed 28 computes ($2.8) Axlotl is what the top fine-tuned LLMs are trained on Deepspeed provides distributed training Flash attention lets data stay on GPU Sample packing packs samples of different lengths into equal length tensors Visualize the RANK of a token in a generated stream instead of logprob The Knowledge Project. Tomorrow Gayner What I’d like in my obituary: Anand was happiness. A guru. Generous. To get what we seek we must deserve this. Build, measure, learn If you did the same thing daily for 50 years, would it be a great thing? If yes, do it. If not, stop. Do this in daily retrospectives My new role should be productivity through technology innovation. That may mean a CTO role. But be specific otherwise no one will understand it Hidden brain podcast. Us 2.0. Win hearts, then minds When in an interaction, ask yourself. Can I learn and change myself? Can I win their hearts, then mines, so their behavior will change. That identity will change Notice when you get emotionally triggered. That’s exactly when you should not get emotionally triggered Try model humility and moral Look for close to people’s identities in our conversations. What are things they like? What does it mean for them? Simply ask. With that understanding of identity, it becomes easier to reframe things in a way they will understand Bard can talk to Gmail and Google Drive! #PREDICTION As automation takes over these mainstream activities, people will take over the niches. Since expertise like knowledge is fractal, there will be many more segments of one in the future and it will be easier to automate clusters of similar abilities. Recommenders and brands will become even more important Stephen Osserman’s Observables have some nice notes. Visualizing partial election results D3 Force Dilemmas: Data Distortion Sandra Becker’s 30 day D3 course

For those in #Singapore and interested in #datavisualization & #llms, I’m talking about Visualizing LLM Hallucinations at SUTD on Thu 8 Feb at 7 pm SGT. This is for a non-technical audience. We’ll visualize the basics of how LLMs work, how they make mistakes, and at least one technique on how to spot these. https://www.meetup.com/data-vis-singapore/events/298902921/ LinkedIn

This week, I learned: Alzhara is one of the VFX companies that worked on Leo’s hyena scene. Their 3D modeling is incredible. Enterprise scenarios leaderboard. Mistral 7B leads. Veda Srinivasan. How does Google manage culture? AMA sessions Manager feedback. Entirely anonymous. Avoid taking feedback for teams less than 5 Workplace concerns team exists. Put managers on watch Books Mohammad Younus. Three zeroes book. Read about his social business theme Pluriverse. Anti fragile. Aurobindo Vedas. Barry Oshry. Seeing systems. Runs workshops but book is better Raghu Anantanarayana has written about Indian archetypes based on Mahabharatha India that is Bharath. Sai Deepak. Podcasts Listen to Nilesh Oak. Sugreeva’s Atlas. Pankaj Tripathi podcast on geography influences acting Areas of focus “I’m an Expert on synthesis and implementation” Intersectionality is another word for complex failures. Also for deep segmentation. Swiss cheese model. Dialogic self theory is about multiple voices in the head. How do we make meaning? Psychological rupture is when cognitive activity is maximum. At any point there are MULTIPLE voices in our heads that are sources of action. We don’t listen to them. Epistemology. Language determines thought. like the word productivity. How does appreciation of a rose become productive? Words from other languages may have incredible power. From other cultures. Paul Sloan. Lateral thinking podcasts from multiple sources Deliberately engage with topics randomly. Deliberately engage with random people Read a random book from the library Watch a random film in a different language Consciously where the six thinking hats or look hard for the silent voices in your head and express them Ask children. They tend to think of more creative and childlike solutions He converted a hiring process into a contest Constantly ask yourself. What if every assumption I’m making about this is wrong? Scenario planning is really about this. List a few scenarios. They’d have high impact or high probability. What happens in this scenario? Ideate You can @mention GPTs to ask a specific GPT a question in ChatGPT. This is really powerful. Hidden brain podcast. Making the most of your mistakes FIX every small mistake. You never know how they might line up in the future You also never know how small little things done well might line up to give you a boost in the future The Toyota cord does not actually stop the production line. It brings a team lead over who quickly diagnoses the problem with you. The responsiveness of the league is a critical factor and so is encouragement That isn’t always a single bottleneck to stop that is the case of a simple failure. There can be a series of holes that happen to align perfectly. These are events that lead to catastrophic failures or successes Do as little as possible, waste as little as possible, until you know that the outcome is worthwhile. Figure out what is the value of the outcome and the most important piece of information you need to discover that Do full research before you try and fail. The aim of failure is learning at the least possible cost How I write podcast. 2023 summary Ask for feedback from friends in a specific way. What 20% should I retain no matter what? What 20% should I cut? This allows them to compliment while providing genuine feedback Hire lawyer interns to proofread. They are the ones that find fault the best Be in a segment of one. Where there is zero competition. Something only you can do Don’t try to do stuff faster. Try to do stuff you don’t want to stop doing Read books older than 50 years Read Michael Collins book on things that sustain Temp service make sure he has some energy to spare. Cuz Riley does the opposite. She waits till she can’t stand it anymore and then writes like crazy until she drops dead. The former leads to thoughtful writing. The latter is emotionally powerful. Be able to do that Vanna is a SQL generation LLM. An alternative to SQLCoder. This thread has a detailed discussion on SQL generation and BI Intel developer cloud has a liberal GPU in the free tier. OpenAI releases text-embedding-3-large which can be truncated. The embedding values have descending importance, so picking the first n is a good approximation. Also, gpt-3.5-turbo-0125 is 50% cheaper. AppAgent is an LLM that can navigate mobile / web apps Retrieval Centric Generation is an emerging alternative to RAG, where the LLM is explicitly built to leverage external knowledge. SimplyRetrieve is an early implementation. Big Code Models Leaderboard is a leaderboard for open source code models.

text-embedding-ada-002 used to give high cosine similarity between texts. I used to consider 85% a reasonable threshold for similarity. I almost never got a similarity less than 50%. text-embedding-3-small and text-embedding-3-large give much lower cosine similarities between texts. For example, take these 5 words: “apple”, “orange”, “Facebook”, “Jamaica”, “Australia”. Here is the similarity between every pair of words across the 3 models: For our words, new text-embedding-3-* models have an average similarity of ~43% while the older text-embedding-ada-002 model had ~85%. ...