Hriday039

Multimodality, Continual Learning | AI Consultant | Anime Freak

[Anatoli Kopadze @AnatoliKopadze]

Anthropic engineer:

"You don't need better prompts. You need graph engineering that makes agents remember everything."

In 25 minutes he shows how engineers at Anthropic wire agents into a graph where each has a job, they run in parallel, verify each other, and share a memory that never resets.

This beats any paid course on building agents I've seen.

Watch it, then read the full guide on graph engineering below.

[Roan @RohOnChain]

https://t.co/8HHG7saiut

[ali @waterloo_intern]

https://t.co/aivc7B4A7Z

[Kalyan KS @kalyan_kpl]

It is a must watch playlist for LLM beginners.

This playlist "Understanding Large Language Models" covers

1️⃣ What is temperature in LLM?

2️⃣ Pretraining Large Language Models: Everything You Need to Know!

3️⃣ How Attention Mechanism Works in Transformer Architecture

4️⃣ What Are Word Embeddings?

5️⃣ LLM Training Starts Here: Dataset Preparation & Tokenization Explained!

6️⃣ LLMs Like ChatGPT, Explained Visually – How Do They Really Work?

Refer to the comments for the playlist link.

[Kalyan KS @kalyan_kpl]

Google DeepMind introduced "GenAI Processors" Library

GenAI Processors is a lightweight Python library that enables efficient, parallel content processing.

GenAI Processors library helps to build Modular, Asynchronous, and Composable AI Pipelines for Generative AI.

✨ Key Features

Modular: Breaks down complex tasks into reusable  Processor  and  PartProcessor units

Integrated with GenAI API: Includes ready-to-use processors like GenaiModel for turn-based API calls and LiveProcessor for real-time streaming interactions.

Extensible: Lets you create custom processors by inheriting from base classes or using simple function decorators.

Rich Content Handling: Supports various content types (text, images, audio, custom JSON).

Asynchronous & Concurrent: Built on Python's familiar asyncio framework to orchestrate concurrent tasks (including network I/O and communication with compute-heavy subthreads).

Stream Management: Has utilities for splitting, concatenating, and merging asynchronous streams of ProcessorParts.

Library - https://t.co/AUoDVfHkuS

[Hriday039 @FHriday039]

🚀 Say hello to Kothok – the next-gen AI conversational assistant by Visie! 🌟
Not just a chatbot, it’s a tailored solution to streamline tasks, analyze your data, & transform customer interactions.
#ConversationalAI #DigitalTransformation #SmartBusiness #KothokAI #VisieTech

[VISIE @visieai]

Enhance customer engagement and boost sales with Kothok, the AI chatbot designed for retail. Provide instant support, personalized recommendations, and seamless shopping experiences.

#AIchatbot #RetailTech #CustomerService #visieaitech

[Hriday039 @FHriday039]

"Had an amazing time conducting the seminar 'Empowering Future Data Scientists' at Daffodil International University! 🌟 The participants were curious, energetic, and asked such insightful questions—truly inspiring. 👏💡

#DataScience #FutureLeaders #DIU #Mentorship #visie

[Asif Shaikat @asifshaikat]

Much needed eye opening piece

The Observability CAP Theorem https://t.co/Vkh8FUKWIg

[Antaripa Saha @doesdatmaksense]

came across a gem of a channel for learning gpu programming, covers
> kernel basics
> code simple neural network by just using cuda
> cuda memory
> performance factors in GPU

wonder why these type channels and videos are so underappreciated

[GeniusThinking @GeniusGTX]

This is Peter Thiel.

He's worth $7.1 billion after founding & investing in $1B startups.

From creating the "PayPal mafia" to hiring Sam Altman, Reid Hoffman, and partnered with Elon Musk...

He is also the greatest recruiter ever!

His 4 questions to spot future billionaire:

[oggy @oggyxe]

if you guys are interested in how the brain / memory works and how it all connects with neural networks, definitely check this guy. immensely based

[Hriday039 @FHriday039]

Nowadays, it's alarming to try to solve every problem with machine learning/AI. But I think we need to about when not to use ml/AI as well...
What's your thought?

#ml #ai #randomthought #mlmisuse #whennottouseml

[Jacob Rintamaki @jacobrintamaki]

So previously I had made a video explaining transformers from the atom up that had some technical inaccuracies. So I fixed them.

Here's the new video and longer thoughts below:
https://t.co/IHCIotGneb

[Gabriel Guerra @gabe_guerra_]

Reading recommendation for this week - A blog detailing an architecture that uses Firecracker VMs for a new type of reconciliation loop. Unlike Kubernetes, this approach recovers not just the application, but also its state.

[Jeff Schneider @jeffrschneider]

Four things caught my attention last week.

In no particular order:
1. @AnthropicAI released artifacts
Mostly hate it.
We implemented something similar in our chat app last year. The idea is to use a system prompt to catch specific requests to build tiny apps (like in html). It returns the source as part of output stream with a tag that the client UI knows about and can render. It's fun - it's cool - but it's a truly awful way to write even a small application and noobs will do it anyway. It's the perl of LLM scripting.

2. @dharmesh released email agents.
Love it.
The simple idea: call agents by sending an email, and receive the response in a return email. This was a but of a 'duh' moment for me. SMTP acts as a ubiquitous decoupling device and is easy for an LLM to parse.

3. @SierraPlatform released an agent benchmark.
Good Start.
I applaud anyone that releases open benchmarks for assessing the utility of AI / agents. This looks like a good start. Their v1 of the agents is simply tools w/function calling (plain jane), so most of the initial results are really bad. It's all Python and the function signatures were not done w/annotations. Need to think about that.

4. @SullyOmarr released constrained tabular agents.
Love it.
Unlike many other agent companies, Sully released a UX model that is tabular (spreadsheet-ish) and offers a set of Tools (or agents) that run across the data. It's a great example of how to ship something useful in a limited amount of time.

[Virat Singh @virattt]

I finally looked at the Apple Intelligence architecture today.

Main thing that stands out is the orchestration of multiple models.

Especially between on-device and server models.

1 • Small LM (SLM) that is 3B param and runs on your device. Vocab size of 49K.

2 • Large LM (LLM) that runs on the server and handles complex tasks. 100K vocab.

3 • Diffusion models for all image generation needs.

4 • Adapter layer to route multimodal user requests to server model.

For evaluation, Apple used human graders to compare performance.

The 3B param on-device SLM outperformed 7B mistral.

The server LLM beat gpt-3.5 and mixtral 8x22B, but fell short of gpt-4.

Excited to see a push for smaller, on-device LMs from the big players.

[Caglar @caglar_ee]

Video lectures, Harvard University CS 229br Advanced Topics in the theory of machine learning spring 2021, by Boaz Barak
https://t.co/jQoKCuEMB3

[Caglar @caglar_ee]

Video lectures, Caltech CS 159 Advanced Topics in Machine Learning Spring 2021 by Yisong Yue, Ugo Rosolia & Jeremy Bernstein
https://t.co/f80OoFVfsG

[sentient being @lwasinam]

Sensei @karpathy is cooking

[Alex Yanko 🇺🇦 @LeopolisDream]

Welcome the new architecture:

Terminator

No residuals, no dot product attention, no normalization...

https://t.co/ZelcVNMkas