Jason Rich Darmawan

Large Language Model Researcher |  Apple Developer Academy Indonesia Alumni

[Adekunle @Gotimehomie]

This was one of the most heartbreaking moments in the entire show whoever wrote this episode needs their writing privileges revoked.

[Jason Rich Darmawan @jasonrichdarma]

@fkysly 租用按需 GPU 吧。DGX Spark 的 L2 缓存和带��太小

[Jason Rich Darmawan @jasonrichdarma]

@natalieyeo @irvinxyz The resources are too finite to do slop hardware. Imagine if we discover AI when we are a space faring civilization. A 17 year old might make a factory in planet Kepler and we get to watch it here instead of the simulation🤣

[Jason Rich Darmawan @jasonrichdarma]

@jes5199 Best guess: OP might found a planet nobody know existed. My research can generate hundreds of GBs of data. I research the knowledge inside an AI. Now imagine NASA recording the universe. The data is not in TB but undefined because the space always expands.

[Jason Rich Darmawan @jasonrichdarma]

@XiaojieLiu1993 @plazined Frontier has millions or at least thousands to burn. It’s a money game now. A short paper costs at least 500 RMB to rent 4x 3090 for 5 days. Each 3090 is to benchmark methods. i.e., 500 RMB is just to fill 1 subsection of the paper and doesn’t include trial and error cost, etc.

[Jason Rich Darmawan @jasonrichdarma]

@GuillaumeLample Le chonk!

[Jason Rich Darmawan @jasonrichdarma]

@IanL556 @0xCE7E5 It depends on the training tokens per parameter and stage. We don't know both your and OP tokens/parameter; or is this fine-tuning or pre-training.
For context, for pretraining, 3773 tokens/parameter (0.026%) is minimum; for fine-tuning 0.7 tokens/parameter (70%) is minimum.

[Konstantin Saifoulline @konstantinsaifo]

Claude Opus 5.5 designed a robot dog from scratch. Every part, every wire, the path planner.

618 parts. 62 cable runs, each one going to a real connector.

Watch what it does when a crate drops in its way.

[Jason Rich Darmawan @jasonrichdarma]

@schneidenbach I abandon the idea to buy a DGX Spark in March 2027, right after graduation. I am going to do PhD, lab will provide GPUs.

[Jason Rich Darmawan @jasonrichdarma]

@schneidenbach This price hike made me double check. 1) have access to rent $0.2/hr/3090. Even though I spent $19/day/4x 3090, it’s not every week without stop. 2) DGX Spark only has 24 MB L2 cache, forcing to use smaller batches or risk multiple GPU<->VRAM round trip for pp/batch

[Historic Vids @historyinmemes]

Dorothy Vaughan adapted from a NASA human computer to a programmer, mastering IBM technology and helping her colleagues transition with her.

Technology has made great strides since then, making quite literally anything possible for companies and everyday users of AI.

[Jason Rich Darmawan @jasonrichdarma]

@KieranLeeDoyle @chubock @MovieplugTV The key nuance is “if we disagree, then there might be a variable that we have not realized yet”. In this case, the host will only open the door with a goat.

[Jason Rich Darmawan @jasonrichdarma]

@InterCeptorLOL @tomshardware AMD dethroned Intel. I don’t know what Intel did wrong except alienate the smartphone chips market and do premium pricing. NVIDIA did alienate gaming and do premium pricing, but NVIDIA moat is CUDA, extensively used by students. AMD needs miracle to dethrone NVIDIA.

[Jason Rich Darmawan @jasonrichdarma]

@KieranLeeDoyle @MovieplugTV I think the correct idea is “population probability”. One of the two unopened doors have 2/3 chance. We just don’t know which one.

I disagree with the idea of “transfer the entire remaining 2/3 onto the single unopened door left on the table”. We can keep our choice and get 2/3

[Jason Rich Darmawan @jasonrichdarma]

@chubock @KieranLeeDoyle @MovieplugTV Agree. The idea of transferring the 33.3% probability to the next door doesn’t make sense. The probability from the opened door split evenly to both unopened doors.

Population wise, one of the door do have 2/3 chance now. We just don’t know it.

[Jason Rich Darmawan @jasonrichdarma]

@NVIDIAGeForce #RTXON

[NVIDIA GeForce @NVIDIAGeForce]

We're giving away a CONTROL Resonant custom wrapped NVIDIA GeForce RTX 5080 to celebrate its official release with #RTXON.

To enter:
🟢 Share this post
🟢 Comment #RTXON

T&Cs: https://t.co/Mbf7w3j1q3

[Jason Rich Darmawan @jasonrichdarma]

@_sanjoydas Additionally, moving info in stages allows the model to process specific information with the same MLP weights. This advantage is shown in the small parallel model capability to generalize knowledge better than small sequential model.

[Jason Rich Darmawan @jasonrichdarma]

@_sanjoydas Sequential model: Small model move info to the last token position immediately. Large model do not do this.
Parallel model: Both small and large move info to the last token position in stages.
In short, you need attention module in each layer to progressively move information.

[Jason Rich Darmawan @jasonrichdarma]

@KieranLeeDoyle @MovieplugTV Ah I got the logic. Thank you.
- “There is 2/3 chance the car is behind one of the two unchosen doors”
- “Transfers the entire remaining 2/3 onto the single unopened door left on the table”