[March 24] Interesting Things I Learnt This Week

 1. Ultrasound becoming Utra small - How MEMS technology has miniaturized ultrasound imaging. It discusses the limitations of traditional ultrasound machines. These machines are bulky and require multiple probes. MEMS technology allows for the creation of a single probe that can image the entire body. This probe is small enough to fit in a lab coat pocket.  The article also details the technical aspects of MEMS transducers.

My Take: Any and all technology is bound to get cheap and miniaturized over time with progress in science and technology. Utrasound is something which can be thought of a speaker and a mic in very layman terms. These are bound to get smaller over time and be more effective. But like any other tool, this can be misused and the reason for skewed male/female ratio in India can also be attributed to ultrasound. My worry is if this tech is available more widely and portable, it will further skew the gender ratios in India(and probably a lot of other countries). This particular tech has to be introduced with great care and all the different social impacts be studied carefully so that proper safeguards are introduced before it becomes widely available. 

 

2. Reverse Engineering Perplexity -  Details about how perplexity is probably working. Not a definitive one but a very good speculation. It seems like perplexity basically summarizes the content from the top 5-10 results of google search. If you search for the exact same thing on google and perplexity and compare the sources, they match 1:1. 

My Take: This is a very interesting take. Some others have speculated that they used Bing instead of Google. We can probably build a browser extension which can do this for any search engine. You don't really need billions of dollars except for the indexing those 4 links. I may attempt it sometime, will let you know.


3. OpenAI GPT-4 vs Groq Mixtral-8x7b - This is an article about comparing OpenAI GPT-4 and Groq Mixtral-8x7b. It discusses parsing speed and accuracy. OpenAI GPT-4 is more accurate but slower. Groq Mixtral-8x7b is faster but makes some mistakes. The author thinks Mixtral-8x7b could be improved with better prompts.

My Take: I think this article is more worrying for NVIDIA than its for OpenAI. Their hardware specialty got a very good challenge from Groq LPU. Models will eventually reach the statis like state which the databases have and with so many open source models out there, it will be a challenge to keep these closed sources models relevant especially if you have to deploy them on laptops, mobile phones etc. Groq is very promising and I hope they can break into the hardware world. Then there is TPUs from Google which have remained deployed only in GCP. I hope Groq's LPUs does not end up like that.

 

Comments

Post a Comment

Popular posts from this blog

My learnings at Google

I just quit Google after almost 7½ years. When I had joined Google in 2016 I was exploring to start up with a travel planning, sharing and booking platform. But then when I got the chance to join Google I told myself that I will stay for 3-4 years and learn how Google's culture makes it build products that people love, at scale. Also, their engineering culture was something I had heard a lot about and thought this is my best chance to learn about and implement it in my startup. But then I managed to overstay at Google for various reasons.  There are various posts talking about what is wrong about Google here , here and here . But I wanted to look at its positives and I hope companies adopt it rather than a toxic culture which is brewing everywhere. Here are a few things which stood out for me (in no particular order): 1. Blameless postmortems : Blameless postmortems are a powerful tool for learning from incidents and improving system reliability. The key aspects of it are   a. Foc
Read more

Top reasons why Tata Elxsi should be blacklisted

Well that a provoking first line, but then thats exactly what I have to say about this company. Tata Elxsi should be blacklisted, but wait a minute by whom? Well the software developers fraternity. Its a blood sucking corporation which is right documented in the documentary The Corporation . Ok why am I telling this after almost one year of leaving it. Well when the fact is that I have nothing to do with it. I have already let out my personal experience here , here and here . Well it was not just me as its obvious from the comments but then it was general feeling of my friends there. Well all I can talk about is friends because I dont know how each and every employee of Tata Elxsi felt. Anyways heres one more story about the kind of harassment and employee of Tata Elxsi might have to suffer . Though I had read it earlier but am blogging now about solely because I wanted to confirm the news being true. And sadly it is true. An employee was indeed made to resign. And indeed its legally
Read more

[14th April 2024] Interesting Things I Learnt This Week

Image
  1. Opera adds built-in support for local LLMs - Opera is adding local AI models to its browser. Users can choose from 150 local LLM variants, keeping their data private on their device. This is part of Opera’s AI Feature Drops Program for early adopters. My Take : Locally running LLMs will become a commonplace very soon. It requires not just software but hardware support as well. On older hardware it might takes ages to get anything done. But on newer hardware it will be good I guess. The best software to get them integrated is going to be the browser for most folks IMHO . I was hoping Firefox to be the first one to do it, but I wish they work towards it. I have no hope of any of the big personal computing softwares to be doing it. They will push for only their models to be running and working across devices, also trusting them to not harvest data off it, would be a challenge. Never the less, integrating LLM APIs in browser will open up some interesting avenues for web applications.
Read more