Hello everyone, this is Kurian posting about some interesting things which happend in the AI world in last week.
This week everyone was working so hard. Lot of AI news to cover this week with almost 2+ news daily.
Why are you all working so hard to make my job so difficult when I write content about things which happend in the last Week in AI world :P
It’s crazy as my friend Shahul mentioned the irony that we invented AI to make life easier, but all the tech folks are working more hours than before due to the same.
Do checkout this weeks news and if you find it interesting do let me know via comments or if we ever meet IRL. Let’s go onto the news now:
🗓️ Monday:
1️⃣ Netflix cautions against blindly using cosine similarity as a measure of semantic similarity between learned embeddings, as it can yield arbitrary and meaningless results.
2️⃣ Some interesting thoughts by Peter Gostev on how LLMs are a lot more cheaper than previous tradional NLP techniques.
3️⃣ Researchers have introduced a ‘Mind Wipe’ technique for erasing hazardous knowledge from AI systems, ensuring functionality remains while enhancing safety. Alongside, the Weapons of Mass Destruction Proxy (#WMDP) benchmark, with 4,157 questions targeting biosecurity, cybersecurity, and chemical security, has been made public.
🗓️ Tuesday:
1️⃣ Infrastructure details for training llama3 models by facebook has been released.
2️⃣ OpenAI team released something open-source in a while. Transformer Debugger(TDB) is a tool developed by OpenAI’s Superalignment team with the goal of supporting investigations into specific behaviors of small language models.
3️⃣ Devin AI, the first AI software engineer was really the news of this week. Let’s even cover reaction of this news separately.
4️⃣ code2prompt, a CLI tool to convert your codebase into a single LLM prompt with source tree, prompt templating, and token counting was released as open-source software with MIT License by Mufeed V H
5️⃣ Santhosh Thottingal gave a fabulous talk on AI and making it work in my mother tongue Malayalam. The youtube video was published on this day while it was actually delivered in a National Seminar organized by the Tirur regional centre of Sree Sankaracharya University of Sanskrit on January 6, 2024.
🗓️ Wednesday:
1️⃣ I compiled the reactions to the news of Devin AI, the first AI Engineer by various folks.
2️⃣ Claude 3 family of Haiku models was released. Haiku is the fastest and most affordable model in its intelligence class was released by Anthropic.
3️⃣ Modular with their Max Engine’s can give 2-5X improvement without any quanitzation or tricks which reduce the accuracy.
🗓️ Thursday:
1️⃣ AI4Bharat team released Indic LLM Suite, a blueprint for training and fine-tuning LLMs in Indic Languages.
2️⃣ Hrishi Olickel who is the CTO of Greywing has been writing some awesome articles in huggingface community blog about how to make better RAGs(Retrieval Augmentation Generation). Do check his articles:
3️⃣ Chip Huyen went through most popular AI repositories in github, categorized them, and studied their growth trajectories. Check the full analysis in blog and tweet.
🗓️ Friday:
1️⃣ Last week, I mentioned about ragas by Jithin James and Shahul ES (my class-mates as well) being selected for Y Combinator. This was featured in one of leading news dailies in Kerala, Mathrubhumi. I appreciate their editor Manoj K Das and R Roshan for featuring them in your esteemed news daily.
2️⃣ Pratik Desai founder of KissanAI announed a new series of fine-tuned Vision LLMs for pest and disease detection and conversation over cure, symptoms, severity and prevention. The Dhenu-vision-lora-0.1 is fine-tuned Qwen-VL-chat, for 3 major crops and 10 diseases, giving 2x performance boost over the base and was trained on synthetic data generated for around 9000 disease images.
3️⃣ Govt of India released an updated advisory toning down what they said earlier. The advisory has been sent only to 8 large social media like organization, some upcoming well-funded AI startups in India has been exempted from this for now.
4️⃣ Hiring managers are now expecting like 6+ years of experience in GenAI, this reminds me of one post by creator of FASTAPI Sebastián Ramírez who said even he didn’t have 5+ years of experience in FASTAPI when someone asked for that when hiring.
5️⃣ Google released Cappy, a small pre-trained scorer model that enhances and surpasses the performance of large multi-task language models. Cappy has been tested across a variety of complex tasks from PromptSource and Big-Bench.
🗓️ Saturday:
1️⃣ Apple announces the paper MM1 - Methods, Analysis & Insights from Multimodal LLM Pre-training. They discuss how to build good performance multi-modal LLMs which means LLMs would be able to see, speak etc. in future beyond what it’s now already doing well that is reading and writing.
2️⃣ Shaheen Gemma 7B, a model being finetuned on Urdu Alpaca dataset. It’s great to see more fine-tuned LLMs in all regional languages in India. Lot of folks are putting effort in bringing my mothertongue language to fore in realm of Generative models.
3️⃣ Anwesha Sen wrote a very well written blog post about the previous AI advisory by govt of India and talk about it’s vague clauses, terms which was like a stepping back into license raj.
🗓️ Sunday:
xAI open sourced their chatbot Grok by releasingg the weights and architecture of our 314 billion parameter Mixture-of-Experts model, Grok-1.
Writing this article and compiling these news took me about 5+ hours. So your contribution to help my work by buying me a coffee would be absolutely pixel-perfect amazing.