Tamaghna Dutta

Why is Cyber Crime @Cyberdost debiting money from my account? And blocking my account without my consent?

@IDFCFIRSTBank - I want to immediately escalate 2 issues as I am unable to get through to customer support since the past 30 mins. - I am unable to do any transactions using UPI with my IDFC First account - I see a debit of INR 12000 from my account w.r.t. some Court Order.

People have been calling AI a Services revolution since #Agents went mainstream, "Service as Software" being the keyword. We've spent the last year building conviction on this thesis and are proud to introduce everyone to @letskomplai

before claude code was a thing, shipping required steep engineering and teams were forced into discipline. they spoke to customers, did their diligence, and stayed lean. because the cost of building the wrong thing was months of wasted work. the friction made a strong filter. but now, teams skip straight to implementation. most teams are living in llm psychotic loops. they build slop, realize it is wrong, let the model suggest a product pivot, scrap everything, and restart. no customer conversations, no market feedback, no testing. just endless slop. infact, multiple teams in the same org are now duplicating efforts because nobody's coordinating anymore. the hard problems remain unsolved because easy problems give you that dopamine hits. everyone's shipping, no one seems to build. i thought cheap execution would democratize creation but it is actually increasing dysfunction.

@theniyo - I am quite frustrated with your customer support. I ordered for Forex cash on Saturday and uploaded all necessary documents. Support told me my documents are actually verified but due to some technical issue it is pending verification on your platform/app.

@akshay_pachaar FSDP and ZeRO Stage 3 (deepspeed) address this with some differences in methodology for both. Both do all-reduce before forward pass and reduce-scatter after backward pass. Former does it at a nn.Module level and latter does it at a layer level. Lovely breakdown though! 🙌🏽

By default, deep learning models only utilize a single GPU for training, even if multiple GPUs are available. An ideal way to train models is to distribute the training workload across multiple GPUs. The graphic depicts four strategies for multi-GPU training👇

You’re in an ML Engineer interview at Google. Interviewer: We need to train an LLM across 1,000 GPUs. How would you make sure all GPUs share what they learn? You: Use a central parameter server to aggregate and redistribute the weights. Interview over. Here’s what you missed:

pantheon is such a good show!

Scaling up RL is all the rage right now, I had a chat with a friend about it yesterday. I'm fairly certain RL will continue to yield more intermediate gains, but I also don't expect it to be the full story. RL is basically "hey this happened to go well (/poorly), let me slightly increase (/decrease) the probability of every action I took for the future". You get a lot more leverage from verifier functions than explicit supervision, this is great. But first, it looks suspicious asymptotically - once the tasks grow to be minutes/hours of interaction long, you're really going to do all that work just to learn a single scalar outcome at the very end, to directly weight the gradient? Beyond asymptotics and second, this doesn't feel like the human mechanism of improvement for majority of intelligence tasks. There's significantly more bits of supervision we extract per rollout via a review/reflect stage along the lines of "what went well? what didn't go so well? what should I try next time?" etc. and the lessons from this stage feel explicit, like a new string to be added to the system prompt for the future, optionally to be distilled into weights (/intuition) later a bit like sleep. In English, we say something becomes "second nature" via this process, and we're missing learning paradigms like this. The new Memory feature is maybe a primordial version of this in ChatGPT, though it is only used for customization not problem solving. Notice that there is no equivalent of this for e.g. Atari RL because there are no LLMs and no in-context learning in those domains. Example algorithm: given a task, do a few rollouts, stuff them all into one context window (along with the reward in each case), use a meta-prompt to review/reflect on what went well or not to obtain string "lesson", to be added to system prompt (or more generally modify the current lessons database). Many blanks to fill in, many tweaks possible, not obvious. Example of lesson: we know LLMs can't super easily see letters due to tokenization and can't super easily count inside the residual stream, hence 'r' in 'strawberry' being famously difficult. Claude system prompt had a "quick fix" patch - a string was added along the lines of "If the user asks you to count letters, first separate them by commas and increment an explicit counter each time and do the task like that". This string is the "lesson", explicitly instructing the model how to complete the counting task, except the question is how this might fall out from agentic practice, instead of it being hard-coded by an engineer, how can this be generalized, and how lessons can be distilled over time to not bloat context windows indefinitely. TLDR: RL will lead to more gains because when done well, it is a lot more leveraged, bitter-lesson-pilled, and superior to SFT. It doesn't feel like the full story, especially as rollout lengths continue to expand. There are more S curves to find beyond, possibly specific to LLMs and without analogues in game/robotics-like environments, which is exciting.

@Saboo_Shubham_ I'd rather use frameworks like LiveKit or Pipecat for Voice agents. Allows for a lot of context engineering, state management, behaviour management, seamless communication with front-end, etc. Do you see any obvious benefits that sets Unmute apart?

Turn any LLM into a real-time voice AI agent. Unmute wraps a LLM into speech-to-text and text-to-speech models, while preserving the reasoning and tool calling capabilities. 100% Opensource.

@socialepfo @PMOIndia I am unable to access the EPFO website on multiple occasions. I want to check my passbook but the website is down or unresponsive most of the time. I am also unable to see the same from Umang as well.

@metaaxiom @PMOIndia Dear member, kindly provide your query.

@socialepfo @LabourMinistry @OfficeOf_MM @mansukhmandviya @CLCMOLE Why is the EPFO website almost never up? Why is there no investment to update the infrastructure and scalability of EPFO's servers? It's always a very frustrating experience.

(50/50) The blend of history, math, and art behind them is as fascinating as it is impactful. Follow for more deep dives! #VoiceSynthesis #ControlTheory #MathInAI #VoiceAI

(49/50) State space models provide a powerful, intuitive framework for understanding and synthesizing human speech.

(1/50) Ever wondered how we mathematically simulate natural human speech? Let’s embark on a journey into state space models—an essential tool for dynamic voice synthesis and control theory.