Mike Hong Profile Banner
Mike Hong Profile
Mike Hong

@MichaelJHong

6,142
Followers
3,477
Following
30
Media
135
Statuses

I build AI applications for B2B. Founder at @evioslab . Partner @chilantro .

Austin, TX
Joined March 2011
Don't wanna be here? Send us removal request.
Pinned Tweet
@MichaelJHong
Mike Hong
5 months
How to convince your wife in the middle of cold winter why "we" should buy 2 x Nvidia 4090s: 2 Nvidia 4090s x 400 Watts = 800 Watts x 3600 seconds (per hour) = 2,880,000 Watts-seconds / 1054 = 2,739 BTUs per hour We can keep our heater off at our study :-)
19
10
251
@MichaelJHong
Mike Hong
3 months
Once upon a time… GPU was feeling a bit fragmented & needed some parallel processing, so it went to Therapy Processing Unit (TPU) for tensor training. Then it met IPU at therapy and was like… whoa, I thought I live in the Matrix… Feeling a bit less fragmented, GPU called
2
1
14
@MichaelJHong
Mike Hong
7 months
e/acc - which one of you geniuses came up with it?
3
6
107
@MichaelJHong
Mike Hong
6 months
Words have so much meaning in our lives. It can inspire us, make us feel loved and uplift us. It can also demotivate us, exacerbate fear, and alienate us. These are the words I've observed in AI headlines just from today: doom death cannot weapons strikes urgent ban concerns
4
9
82
@MichaelJHong
Mike Hong
6 months
Someone just told me that global data consumption & creation will hit 1 Yottabytes in 3 yrs with AI/ML model development in the future. To put that into context, a single yottabyte could stream high definition (1GB/hr.) video continuously for over 114,100 years, about 1,426
Tweet media one
Tweet media two
11
7
74
@MichaelJHong
Mike Hong
6 months
Reduce spam by adding "+" after your email when signing up to things, e.g. yourname+freetrial @NewCo .com If "freetrial" later appears from a different sender, the original site likely sold your address. You can reach out politely to "freetrial" to remove you.
Tweet media one
3
3
73
@MichaelJHong
Mike Hong
2 months
Why don’t we integrate the backdrop technique directly into the chunker? Anyone built something like this?
0
0
13
@MichaelJHong
Mike Hong
7 months
Leaked GPT-4 pricing is fake. But… wouldn’t it be nice?
Tweet media one
0
1
40
@MichaelJHong
Mike Hong
6 months
“Apache Airflow is a work flow management. Actually, it is more like a job scheduler on steroids. When it comes to ETL, it relies on other tools and scripts to do the heavy lifting, it’s like a party planner that brings all the right people together for a grand event, but it
2
1
55
@MichaelJHong
Mike Hong
6 months
If we can get a system under $20K capable of handling payloads over 30 lbs... I think we are ready for cost effective actuators and end effectors with higher torque & payload. Kuka & Fanuc, are you ready to tier-down now?
@zipengfu
Zipeng Fu
6 months
Introduce 𝐌𝐨𝐛𝐢𝐥𝐞 𝐀𝐋𝐎𝐇𝐀🏄 -- Learning! With 50 demos, our robot can autonomously complete complex mobile manipulation tasks: - cook and serve shrimp🦐 - call and take elevator🛗 - store a 3Ibs pot to a two-door cabinet Open-sourced! Co-led @tonyzzhao , @chelseabfinn
187
885
4K
3
6
49
@MichaelJHong
Mike Hong
6 months
When should you use or not use GraphQL?
7
0
61
@MichaelJHong
Mike Hong
7 months
my X feed blowing up with AI influencers.
Tweet media one
1
3
60
@MichaelJHong
Mike Hong
7 months
When something is "AI-powered", how do you know if it's actually using AI? Here's a quick stress test.
Tweet media one
4
5
48
@MichaelJHong
Mike Hong
1 year
How to validate whether if someone is actually using AI.
Tweet media one
0
0
1
@MichaelJHong
Mike Hong
1 month
To fine-tune or not to fine-tune, that is the question: Whether 'tis nobler to prove thy worth, before tweaking and adjusting, or to take a model as it is, and risk the slings and arrows of outrageous errors... To prove, to test, to validate, and then to fine-tune, ay, there's
1
1
7
@MichaelJHong
Mike Hong
7 months
Breakfast & coffee. Data. Data. Data. Data.
Tweet media one
1
0
41
@MichaelJHong
Mike Hong
7 months
What if every digital content, i.e. posts on X carried an intelligent digital annotator?
Tweet media one
4
3
42
@MichaelJHong
Mike Hong
6 months
Is anyone working on quantum computers? Apple's work on LLM in a flash is pretty cool near-term. But what's really exciting for me is quantum computing and how it would impact AI ops in the next 10 - 20yrs. Imagine computers
Tweet media one
1
1
34
@MichaelJHong
Mike Hong
2 months
Any recommendations for serverless GPUs? We’ve tried - Banana BaseTen Replicate Inferless Cloudflare Modal We need scale & ease of use…
5
1
24
@MichaelJHong
Mike Hong
7 months
The egg test on Grok.
Tweet media one
0
1
25
@MichaelJHong
Mike Hong
7 months
122 years of recorded history: - Automobiles - Airplanes - Antibiotics - Nuclear energy - Personal computers - Internet - Smartphones - Social Media - EVs & RE - CRISPR - Reusable rockets - Quantum computers 1901 - 2023, only .0049% of the full history of civilization...
Tweet media one
2
0
27
@MichaelJHong
Mike Hong
5 months
guys, don't do this...
3
0
26
@MichaelJHong
Mike Hong
7 months
Just as customers request specific dishes from the menu, clients in a computer network request specific data/services from the server. It can be for web pages, files, data processing, etc... that the server provides. Now.. in 10k zoomed out view, how do restaurants start?
1
0
10
@MichaelJHong
Mike Hong
8 months
@anothercohen Are you sure this wasn't on Sunday?
2
0
16
@MichaelJHong
Mike Hong
3 months
2024…. January 🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🤦‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️😴 February 🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🤷‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️💤 March 🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🏃‍♂️🙏🏃‍♂️
1
0
9
@MichaelJHong
Mike Hong
3 months
Then Jim Cramer came along… THE END.
0
0
8
@MichaelJHong
Mike Hong
1 year
We are moving towards the era of scaling synthetic data ahead of semi-conductors.
Tweet media one
0
0
7
@MichaelJHong
Mike Hong
8 months
@anothercohen oh... they close on Sundays...
5
0
7
@MichaelJHong
Mike Hong
7 months
I agree. Building “enduring” AI applications requires industry-specific, vertical approach. I think the concept of TAM is not static, it’s dynamic. I’ve rarely seen successful businesses that execute first with TAM. They generally start with SOM/PMF, SAM & work their way towards
@christinkers
christine kim
7 months
We are at the early innings of an incredible period of innovation. I believe vertical specialization and domain-tuned AI will be key to building enduring applications. Here are takeaways from my post on why the moment for Vertical AI is now.
17
37
213
0
1
7
@MichaelJHong
Mike Hong
3 months
HPU always talked about how he always knew HPUs that knew QPUs that knew RDUs.
1
0
6
@MichaelJHong
Mike Hong
7 months
To sum it all up. There is a clear similarity that resonates across every sector, businesses & industries. It's always about customers, people and quality of service or product. It centers around - "Who are we here to serve?"
0
0
5
@MichaelJHong
Mike Hong
1 year
It’d be nice to see all the branches converge at one point. Anyone working on routers?
Tweet media one
2
1
4
@MichaelJHong
Mike Hong
1 year
1. GPT-3.5 & 4 2. Falcon 40-B 3. MPT-7B 4. Cerebras-GPT 3. RedPajama-INCITE 4. BLOOM 5. Dolly 6. Flan-T5 7. StableLM-Base 8. Camel
1
0
4
@MichaelJHong
Mike Hong
7 months
What is the statistical likelihood of cybertruck actually needing bullet proof doors?
6
0
4
@MichaelJHong
Mike Hong
1 year
There are so many articles and posts about developing a “perfect” prompt. There is no such thing as a “perfect” prompt. What’s important is having a process to get the output you desire for your own specific use case.
0
1
3
@MichaelJHong
Mike Hong
7 months
I found a really cool mushroom while hiking. Can I eat this?
Tweet media one
1
1
2
@MichaelJHong
Mike Hong
7 months
Can it reason? If it's looking for patterns in massive amounts of data and using these patterns to make decisions + learning from it, it may be using AI.
Tweet media one
0
0
3
@MichaelJHong
Mike Hong
7 months
This an example of AI companies’ liability and risk. An app that could prove useful but also create a tremendous liability.
0
0
3
@MichaelJHong
Mike Hong
1 year
LLMs not commercially licensed: 1. LLaMa 2. Alpaca 3. Vicuna 4. Galactica 5. Koala LLMs that are available for commercial use:
1
0
2
@MichaelJHong
Mike Hong
7 months
We can go on and on, but things do not all align and work on first day. We all spend heavy amount of time debugging and optimizing.
1
0
2
@MichaelJHong
Mike Hong
7 months
Can it move? Does it move without help? If it moves based on what it sees and hears and does not move on a programmed path, it may be using AI.
Tweet media one
1
0
2
@MichaelJHong
Mike Hong
6 months
@shashankydv7 TBH, I don’t know. We are poor predictors of future. Synthetic data makes yottabytes not as science fiction-y, but realistically, it’s not just the data consumption/creation. We’ll need breakthroughs in energy efficient storage, materials for chips, etc… I just think that 3 yrs
1
1
2
@MichaelJHong
Mike Hong
1 year
Can you create a pun on AI? GPT-4: Why did the AI go to school? Because it wanted to become a deep “learner”! Bard: What do you call an AI that’s always getting lost? A navigation unit.
Data bias
1
Algorithmic bias
1
Designer bias
2
None of the above
0
0
1
1
@MichaelJHong
Mike Hong
7 months
Then there's UI/UX in restaurants, which is essentially stripped down to "How do customers feel when they walk into the restaurant?" This is akin to the UI/UX in tech. Choice of lighting, music, table, walls.... Choice of colors, layout and the experience....
1
0
1
@MichaelJHong
Mike Hong
1 year
175B to 540B parameters within the course of 2021 - 2022. Anyone have any idea what 2022 - 2023 looks like?
Tweet media one
0
0
1
@MichaelJHong
Mike Hong
1 year
Taking this home with me.
0
0
1
@MichaelJHong
Mike Hong
7 months
@andruyeung Best of luck. Attended one of your meetups in the past, you bring great value to the community.
0
0
1
@MichaelJHong
Mike Hong
1 year
What other LLMs out there?
Tweet media one
1
0
1
@MichaelJHong
Mike Hong
6 months
@randix93 Qq, in your experience - with complex app requirements, do you think ingress/egress cost savings outweigh the additional costs of running GraphQL? Trying to assess efficiency, dev ops, costs…
1
0
0
@MichaelJHong
Mike Hong
7 months
@BlasikRandy Yes. Tooltip to display bias, context, etc…
0
0
1
@MichaelJHong
Mike Hong
7 months
@TruttMD Ahh… yes, did not think about the use case outside the US. Thanks for the insight.
0
0
1
@MichaelJHong
Mike Hong
7 months
@BlasikRandy Holiday shoppers on shopping cart rage? 🤔
0
0
1
@MichaelJHong
Mike Hong
7 months
@freegeneralist Glad it brings value!
0
0
1
@MichaelJHong
Mike Hong
7 months
Can it hear? Does it respond in a sensible way that's useful? If it's translating what you say, it may be using speech recognition.
Tweet media one
1
0
1
@MichaelJHong
Mike Hong
7 months
Each functional component in the restaurant from fryer to serving lines, plays an important role. Similarly in tech, every functional component-from database to application servers, and APIs play a vital role.
1
0
0
@MichaelJHong
Mike Hong
1 month
1
0
1
@MichaelJHong
Mike Hong
6 months
@kuzetsa I think resource hurdles can be eventually overcome through innovation. My short-term linear concerns are more or less centered around execution & timeline.
0
0
1
@MichaelJHong
Mike Hong
1 month
@SeanMollet Would you mind elaborating?
1
0
1
@MichaelJHong
Mike Hong
7 months
@gdb 🙌🏻
0
0
1
@MichaelJHong
Mike Hong
1 year
UI improvement.
Tweet media one
0
1
1
@MichaelJHong
Mike Hong
6 months
@_Martin_QL_ Cool. I’m following you for updates. Would love to see what you build in the future!
0
0
1
@MichaelJHong
Mike Hong
1 month
@SeanMollet Got it. Thanks for your insight.
1
0
1
@MichaelJHong
Mike Hong
1 year
@nerddoingsocial 🤦‍♂️
0
0
1
@MichaelJHong
Mike Hong
7 months
0
0
1
@MichaelJHong
Mike Hong
7 months
Can it see? If it can identify what it sees and use computer vision & image processing, it may be using AI.
Tweet media one
1
0
1
@MichaelJHong
Mike Hong
1 year
Lower cost, longer testing without any fine tuning. 1024 tokens to execute 2048 token inference. Instead of adding position embeddings at the bottom stack, add linear bias to each attention score.
0
0
1
@MichaelJHong
Mike Hong
7 months
Speaking of APIs, in restaurants, a similar role is played by Expos (expediters). They facilitate communication between BOH & FOH. Just like APIs enable different systems to interact, Expos ensure that order from Ubereats, online, dine-in are accurately prepared and served.
1
0
0
@MichaelJHong
Mike Hong
6 months
@JustDeeevin Works for other email domains as well. Sometimes, companies narrow down the allowlist to prevent “+” usage.
0
0
1
@MichaelJHong
Mike Hong
8 months
@nearcyan Love this. How are handling the images? Are you scraping the first image of the content?
0
0
0
@MichaelJHong
Mike Hong
1 year
My favorite stack.
Tweet media one
0
0
1
@MichaelJHong
Mike Hong
6 months
@Manitcor Got it. For simple app, It’d be an overkill.
0
0
1
@MichaelJHong
Mike Hong
1 year
0
0
1
@MichaelJHong
Mike Hong
7 months
Can it read? If it reads what you type, respond in a sensible way that's useful and can analyze it for patterns, it may be using NLP.
Tweet media one
1
0
1
@MichaelJHong
Mike Hong
5 months
@mikecane Haha. Yes, she then completely broke down my logic by explaining to me the costs & environment impact of running GPUs all day… They are always right, you know…
1
0
1
@MichaelJHong
Mike Hong
1 year
How to keep up to date on Openai's latest blogs and developments -
0
0
1
@MichaelJHong
Mike Hong
9 months
@nickgraynews Kimchi Fries
0
0
1
@MichaelJHong
Mike Hong
10 months
0
0
1
@MichaelJHong
Mike Hong
13 days
@dotinvoke Replicate is great. Trying to lower latency & cost & cold start.
0
0
1