HengJay

Newbie of iOS development / ML & DL domain
LinkedIn Resume

[ちいかわ💫アニメ火金 @ngnchiikawa]

🚗

[サトウ @sato_cube]

ずっと見ていたいと思いまして。

[Heptabase @Heptabase]

Today, we’ve made some big changes to Heptabase.

Many of you use Heptabase as your core knowledge base, storing not just notes but also textbooks, research papers, video lectures, and daily journals. Over the past few months, we’ve seen a growing wave of requests for a NotebookLM-like experience right inside Heptabase—one where you can ask anything, and the AI finds and reads your sources to craft trustworthy responses with accurate citations.

We're happy to announce that we’ve officially brought this experience to Heptabase—and we’ve even gone beyond it. We’ve fully integrated AI into our built-in note-taking, whiteboarding, PDF reading, and video-watching workflows. This means you can not only ask AI about anything in your knowledge base, but also treat it as a reading partner that explains whatever you’re currently looking at. We’ve implemented this in a way that respects user privacy by self-hosting all data parsing and embedding services. No AI provider can see your data unless you choose to chat with them, and you can always specify what AI can or can’t read.

Here are the three major things we’ve shipped in this latest update:

First, we’ve introduced a “Research a topic” button on the left sidebar. This opens an interface where you can upload many types of sources: PDFs, YouTube links, documents, images, and more (Web Cards and EPUB support are coming soon!). Once you click “Start Research,” we add them to a new whiteboard as cards, parse PDF contents, and download video transcripts. This allows you to ask AI anything about the whiteboard and get accurate citations pointing back to specific blocks or timestamps in your sources. You can save any AI response as a card with just one click.

Second, we’ve given the AI a suite of tools to use during its thinking process. Our AI can search your entire knowledge base using both keyword and semantic search, perform deep search inside specific PDFs, and intelligently decide to read more pages or look for more notes when needed. This becomes especially powerful when you’re using premium models such as Gemini 3.0 Pro, GPT 5.1, or Claude 4.5 Sonnet. In my own testing, it was able to pinpoint the best content across my 8,000 notes and 300 PDFs and use that information to provide answers with citations to specific pages.

Third, we’re collaborating with Google to give all Pro subscribers and free trial users a monthly AI credit. You can now chat with the most advanced Google Gemini models at no extra cost. If you hit the monthly limit, you can upgrade to the Premium Plan, which gives you 10× the Pro usage limit plus access to all the latest models from Google, OpenAI, and Anthropic. This means you can always use the most advanced possible model inside your knowledge base. While we can’t guarantee how much AI credit we’ll be able to provide free trial users over time, we can guarantee that right now—and over the next few months—will be a special period during which free trial users can access the most AI during their trial. If you haven’t tried Heptabase before, now is the best time.

There are many tools on the market that allow you to ask AI about your sources, but simply asking AI is not enough to establish a deep understanding of a topic. What matters more is what you do before and after asking AI: READING the sources, WRITING about what you’ve learned in your own words, and MAPPING out how different ideas connect.

These fundamental methods for learning and research remain irreplaceable if you truly want to master something. That’s why a great PDF reader, a rich text editor, and a powerful whiteboard matter. While Heptabase already has some of the best features in these categories, we’re now combining the strengths of knowledge management and AI to create the best possible environment for learning and research.

We’ve updated our website slogan to “Master anything you learn. Do your best research with AI.” We’re excited about everything we’ll be shipping in the upcoming months—features we believe will help you understand and create knowledge in ways you never thought possible.

[Andrej Karpathy @karpathy]

We have to take the LLMs to school.

When you open any textbook, you'll see three major types of information:

1. Background information / exposition. The meat of the textbook that explains concepts. As you attend over it, your brain is training on that data. This is equivalent to pretraining, where the model is reading the internet and accumulating background knowledge.

2. Worked problems with solutions. These are concrete examples of how an expert solves problems. They are demonstrations to be imitated. This is equivalent to supervised finetuning, where the model is finetuning on "ideal responses" for an Assistant, written by humans.

3. Practice problems. These are prompts to the student, usually without the solution, but always with the final answer. There are usually many, many of these at the end of each chapter. They are prompting the student to learn by trial & error - they have to try a bunch of stuff to get to the right answer. This is equivalent to reinforcement learning.

We've subjected LLMs to a ton of 1 and 2, but 3 is a nascent, emerging frontier. When we're creating datasets for LLMs, it's no different from writing textbooks for them, with these 3 types of data. They have to read, and they have to practice.

[HengJay @HengJayWang]

@NVIDIAGeForce #GeForceRTX50 Like RTX50 🥰

[Google DeepMind @GoogleDeepMind]

Gemini 2.0 Flash Experimental has the ability to produce native audio in a variety of styles and languages - all from scratch. 🗣️

Here’s how this is different to traditional text-to-speech systems ↓ https://t.co/FRWb3q3KHe

[HengJay @HengJayWang]

@NVIDIAGeForce #HalfLife2RTX 🥳🥳🥳

[ナガノ @ngntrtr]

👈

[ナガノ @ngntrtr]

🪙

[Lisa Su @LisaSu]

10 years ago I had the honor of a lifetime to be named @AMD CEO. It’s been an incredible journey with so many proud moments. Today I want to thank the global @AMD team for all you do. As amazing as the last 10 years have been, the best is yet to come. 🥰

[Notion @NotionHQ]

Introducing Notion Sites—the easiest way to make a website 🌐

We turned Notion’s "Publish" button into a full suite of features, including custom domains. To help you create personal websites, portfolios, travel guides, and more with just one click.

Get started today at https://t.co/dw8YQoWLqC (P.S. Tag us, and show us what you create!)

[HengJay @HengJayWang]

@janechao_dev @PofatTseng 原來波肥大大現在在Meta服務嗎🤩

[HengJay @HengJayWang]

@m_elhoushi @davidwu1116 @AkshatS07 @bilgeacun @bwasti @Ahhegazy77 @BeidiChen @CarolejeanWu @m_elhoushi Hi, is there any update for releasing the code? Exciting to try it! 🥰

[Andrej Karpathy @karpathy]

no. people misunderstand chinchilla.
chinchilla doesn't tell you the point of convergence.
it tells you the point of compute optimality.
if all you care about is perplexity, for every FLOPs compute budget, how big model on how many tokens should you train?
for reasons not fully intuitively understandable, severely under-trained models seem to be compute optimal.
in many practical settings though, this is not what you care about.
what you care about is what is the best possible model at some model size? (e.g. 8B, that is all that i can fit on my GPU or something)
and the best possible model at that size is the one you continue training ~forever.
you're "wasting" flops and you could have had a much stronger, (but bigger) model with those flops.
but you're getting an increasingly stronger model that fits.
and seemingly this continues to be true without too much diminishing returns for a very long time.

[MrBeast @MrBeast]

I’m gonna give 10 random people that repost this and follow me $25,000 for fun (the $250,000 my X video made)

I’ll pick the winners in 72 hours

[HengJay @HengJayWang]

@pointfreeco How to change the setting so that subtitles to show green highlight like yours videos?

[HengJay @HengJayWang]

@eddiekao @eddiekao 請問龍哥新的第二版也會有電子書版本嗎? 實體書快沒地方放了 🤣

[HengJay @HengJayWang]

The second AWS certification (SAA-C03) I earned at 2023 ! 🥳🥳🥳
View my verified achievement from @awscloud.
#awscertified #getawscertifiedchallenge https://t.co/dUQuQAYoJx

[HengJay @HengJayWang]

@awscloud SAA-C03 will be the next one ! 🤓

[HengJay @HengJayWang]

The first AWS Certificate (CLF-C01) I earned at 2023! 🥳
View my verified achievement from @awscloud. https://t.co/HkHexzyYx7