Things I learned

This week, I learned: Grab. Improving last mile delivery in maps. When did people pick up the phone, when should driver be allocated to minimize waiting time, layer on top of OSM. Singapore developers the Sea Lion 7b model Try VLLM with AWQ format. Can do batch inferencing. Needs a good GPU Amex prediction whether they can pay back in 1 year or 18 months. That choice is a business decision. In real time. Precompute individual score and use it as input to another model. Model must be explainable by regulation. Creates decision tree models therefore. Compliance team must agree if I can use a feature. Can’t use gender. Age (in US, Canada);- high age is more risk. Can’t use edu level in the US. Capture information from camera and use LLMs. Like traffic cameras mapping. Explore GIS from video cameras Grab tracks road closures and road accidents and whether a cycle can go on a road vs a bike vs a car All drivers have a front facing camera Drivers report road accidents by pressing a button Amex prices individual loans when selling to a collection agency #TODO buy a bike head camera! Playwright is a browser-based test framework. Supports recording. OpenAI provides logprobs for tokens! This can be a used to create cool visualizations of the likelihood of the each tokens. Github Copilot’s new features makes your entire workspace or a specific file its context. It also auto-writes your commit messages and PR descriptions. Mixtral-8x7b-Instruct “… really does seem to be equivalent in quality to ChatGPT 3.5.” Ref Practical AI podcast Advent of Gen AI is going on. Explore add to tools in data science course. Model validation write a book as an open source to github repository. Easier to evolve and easier to get feedback on.. Explore utterances as a GitHub commenting platform automatically give credits to contributors who have center pull request that was accepted or an issue that was fixed. This encourages contribution Visit book.premai.io ast-grep is a semgrep alternative that focuses on code refactoring rather than security. Comby is another such tool Serply is a Google Search API alternative to Google CSE ⭐ Generate textbooks! ChatGPT is good at generating questions or training datasets. It genuinely creates them rather than replicating from memory. Ref v0.dev creates web pages from code. Example. LIDA from Microsoft is an LLM based data visualization tool.

This week, I learned: Bard supports extensions that include @Gmail – i.e. converse with your email. llama-cpp-python works with other GGUF models like Mistral and allows constrained output - JSON, function calling, etc. Ref 12 Tuning Strategies for RAG Llama Datasets are RAG datasets created mostly using GPT-4. Mostly small datasets. ⭐ Intuitions about large language models Bigger models (70b) are much better at learning from few-shot examples. They really learn. Bigger models will keep getting better! Chain of Thought prompting is a way of providing more compute to complex problems that require more compute Models will show emergent (completely new) behaviors that can’t be predicted from extrapolation. These may not be intentional. CodeAnt.ai is a VS Code plugin to detect code smells, refactor for modularity, to write docstrings and unit tests Anyscale prices the 7b Llama2, Zephyr, Mistral models at 15 cents per 1M tokens. Roughly 1/10th of GPT-3.5 Turbo’s ~$1.5 per 1M tokens Tools to identify personally identifiable information: galactic can use LLMs to detect PII Presidio by Microsoft Sherlock is a generic sematic type matching DL model pii-extractor-llm was trained on Indian names GLiNER is a Lightweight Generalist model for NER Tools to explore ElevenLabs speaks in your voice Cutout Pro removes backgrounds and parts of images Vocal Remover removes vocals from songs CapCut video editor TheBloke’s $35/month Patreon might be one of the least expensive ways to set up quantized LLMs in production. Microsoft released table-transformer to extract tables from PDFs. Sample usage Convert PDF to markdown with marker - an improvement over nougat. JupyterLab has a %%ai magic to use LLMs within notebooks. Ref Telling ChatGPT that the year is 2123 makes it bypass copyright. Ref Meta released SeamlessExpressive which preserves emotions in speech-to-speech translations Unsloth offers faster lower-memory LLM QLoRA finetuning DeepSeek is an open-source high-quality LLM Scalable Extraction of Training Data from (Production) Language Models extracts training data by repeating a token infinitely. SkyPilot lets you run LLMs on any cloud provider. vLLM lets you deploy LLMs with a single command. llamafile lets you run LLMs locally as a single file executable!

This week, I learned: Gwern Branwen says LLMs nudge his “… making heavier use of the languages I don’t know well (Emacs Lisp & Python) since I increasingly trust that an LLM can help me maintain them.” Undetectable.ai checks for AI content. But it had false positives AND negatives in the 5 checks I ran. GPTZero got 2/2 right and seems better at detecting AI content. CoVA scrapes web pages via OCR When coding with LLMs, have SHORT, RELIABLE feedback loops. Ref

This week, I learned: This is an interesting GPT Vision API prompt from Simon Willison: “given this event flyer, create a link to add it to my Google Calendar”. Ref Quote from Jerry Liu: “GPT 4 is really good at complex reasoning”. It’s worth exploring what that means. Quote from Jerry Liu: “RAG is a hack”. It’s engineered, not machine learnt, so it’s suboptimal. We need an ML way of creating the context. Maybe fine tuning can be a way of CREATING the right context. But RAG can handle deterministic stuff like access control. Open AI fine tuning API is not good at memorizing info the way it is exposed. But the Gorilla paper shows that fine tuning can actually memorize well. Learn ML optimization approach - LLMOps. Have an evaluation framework with metrics like weights and biases or tensorboard. Helps figure out where fine tuning helps and where RAG does. Soon, this will become important. Flat indexing of chunks is not the only way to store embeddings. LlamaIndex allows you to create hierarchies that you can traverse for retrieval Agents mimic programming primitives. Switch. While. Call a function. Print. OpenRouter hosts several models and offers them as APIs! Ragas metrics evaluate quality of a RAG pipeline Orca 2 was trained on different reasoning techniques (e.g. step-by-step) and is as good as larger models Embeddings can help just re-rank regular search results. Ref Claude 2 Anthropic has a 200K context window but is still crap. Video-Llava can understand videos too. CoVA scrapes web pages using LLMs and visual information. jsonrepair can fix JSON fairly well. jsonformer wraps HuggingFace models to produce JSON. Ref Google has a model garden with lots of pre-trained and trainable models. Gorilla LLM specializes in APPI calls: Torch Hub, TensorFlow Hub, HuggingFace GPT-4 does not do abstraction at human levels Each of the GPTs / Prompts we create could be like a UNIX command prompt, and become a startup of its own Llava Plus extends LlaVA with pre-trained vision models that make image editing better Ollama runs local LLMs

This week, I learned: XOT - Everything of Thought is a new prompt from Microsoft but I don’t understand it Creating Fine-Tuning datasets WITHOUT inputs Tamil-Llama Voyager plays Minecraft! Langchain supports evaluators. Pydantic is all you need drives towards code = data = text!

This week, I learned: Julius.ai queries structured data. TODO: Explore https://github.com/microsoft/TaskMatrix microsoft/autogen enables multi-agent conversations. Architecture of today’s LLMs is similar to the A16Z architecture Stanford Foundational Model Transparency index was critiqued as misleading vLLM runs HuggingFace transformers models faster. So does DeepSpeed

This week, I learned: List of Gen AI companies disrupting SaaS incumbents: LinkedIn