Abdul

Software Engineer | boxing enthusiast | weight lifting addict | professional startup starter | Working on tech that matters…to me…sometimes

[Abdul @_abdul_h]

The irony…somethings in life are really just full circle 🤦🏾‍♂️

[Abdul @_abdul_h]

Picked up Rich Sutton's RL book in his class for a mere $20 years ago.His TA gave a speech full of foresight. With Q* buzzing, it's like I'm right back there,book in hand.Dusting off those pages years later, who would've thought.Mind its well-loved look;it's seen some serious use

[Abdul @_abdul_h]

Great insight by @GregKamradt , everything you need to know about OpenAi DevDay in tweet! Couldn’t agree more with #1 and #3 on this list.

[Greg Kamradt @GregKamradt]

My Top 10 Favorite DevDay Announcements

(And the ones that didn’t make the cut at the bottom)

I prioritized them based on: immediate & potential impact + interestingness (totally subjective)

#10 - Custom Models-As-A-Service
You can now work with OpenAI researchers to go deep on model building for your specific use case…it’ll just run you $2M-$3M to do so

https://t.co/ymG5ehRX4D

This is an interesting service play for OpenAI. Also a good exercise for domain specific model building.

What I find even more interesting is the opportunity for other agencies to provide fine-tuning help as a service.

#9 - GPT-4 Fine Tuning
It’s well known that fine tuning gpt-3.5 exceeds GPT4 on some tasks. Even if the performance is the same, you get a massive drop in cost and latency.

https://t.co/ZMo4AfiwSS

Fine-Tuning GPT-4 will take this further. We’ll need to reset benchmarks.

#8 - Text-To-Speech API
We already had good text to speech with ElevenLabs and https://t.co/j5h5Oh1Syx so this isn’t a big unlock. I don’t mind consolidating vendors.

Not to mention, OpenAI’s TTS is 10x cheaper than ElevenLabs.

How long do you think it’ll be till they allow custom voices?

#7 - ChatGPT Backed By GPT-4-Turbo
Love it - I use ChatGPT for ad-hoc tasks almost hourly, I’ll take the speed boost here all day long.

This isn’t a huge change, but the frequency of use puts it in my top 10

#6 - GPT-4-Turbo - 128K Token Context Length
This one’s is a bit controversial - On one hand you have the crowd that adores context length. It’s an easy metric to wrap your head around.

But we could already handle long context via chunking and chain strategies. I agree that it’s a nice addition, but a bit over hyped.

Yes you get lower latency with a single API call, but what is the performance hit?

PS: I’m currently pressure testing the 128K context limit with a “needle in a haystack” analysis. I’ll drop a random fact in the middle of a large context window to see if gpt4-128K context can pull it out.

#5 - GPT 3.5 & 4 Price Reduction (2-3x)
So GPT4 is better and it’s cheaper? I like that combo.

As the price reduces you drop the barrier to entry on a few use cases.

For example a use case I heard yesterday from @niallohiggins, if you were in the data monitoring game, passing in a bunch of logs to GPT4 was expensive.

At half the costs, you can now increase your flow of data

#4 - DALL·E 3 API
Love this one - it’s so high on my list because it’s a net new capability for millions of developers.

Get ready to see millions of photos flooding apps near you.

#3 - Assistant GPTs/API
Ok, now we get into the serious part of the list. OpenAI isn’t messing around anymore

Let’s talk about what the Assistant API actually does:

* Code interpreter API
** This means OpenAI will run a bit of code on their end, then give you the result. This is great for longer math and file transformations.
** This was only available via the UI up until now, having it via API means devs can offload a ton of hacks they did to try to get this working before

* File Retrieval
** So assistants can reference files you’ve uploaded
** OpenAI will take care of chunking, embeddings, and retrieval for you - this sounds great on paper right?
** For no-code builders this’ll be a great unlock they don’t have to worry about more
** But for devs, I predict they’ll want more control over the retrieval process then OpenAI will give them.
** Not only that, but it’s super expensive to host files with OpenAI

* Persistent threads
** This is nice and convenient, thread management per user was an annoying issue for a while

This isn't higher on my list because of plugin adoption. I'm super excited to build with it, but I'll be more excited when I see user energy for assistant value.

#2 - ChatGPT Store
You might think this would be #1, but it was dethroned because of my uncertainty around adoption.

Yes, the demo's cool, but haven’t seen the evidence of user-need to call this one game changing yet. If it does work, then it’s the beginning of a massive product opportunity

There are a couple facts to keep in mind:
* - Building marketplaces is extremely hard
* - Plugin adoption didn’t go well
* + Monetary incentives to build apps can drive crazy adoption

Reasons why I like the ChatGPT Store:
* Instead of “look at these 150 ChatGPT Prompts” posts on LinkedIn well now get “look at these 150 GPTs” instead. In theory the latter should provide more value
* Love the creator profiles
* With revenue sharing, not only do developers have incentive to build, but they have an immediate feedback loop to the value they’re providing to users.

The biggest use case of wrapper-apps has been “chat with your data” which is the bread and butter of these assistants

So the big question in everyone’s mind is, “Are wrappers dead?”

My quick answer is: Not yet, it’s not looking good, but they control their own fate

If you want a bellwether of where this is going, follow https://t.co/ijQIlalSUg. They are about as “wrapper” as it gets.

Yet to beat OpenAI, they’ll need to provide ancillary services, build quality products around their wrapper, and better meet the needs of their customers.

The key here is to position yourself so you’re not competing with OpenAI. Play a different game.

#1 - GPT-4V API
Finally, my favorite announcement from Monday.

Lowering latency and cost are nice, but how often do you have a new modality at your service?

After GPT-4V came out in the UI I wrote a big post on how people were using it

https://t.co/XBOd77Tmo0

They were all one shot examples: you give GPT-4 a single photo and it gives you a response.

But now that the API is opened up we are seeing a new class of interactions

People are now demoing real time identification from their webcams and GPT-4V controlling computers

https://t.co/8Mw5DTFzX8
https://t.co/nm0QUu2S2w

This opens up semantic search and understanding of video - huge applications for interpreting:
* Computer screens
* Security cameras
* Meta’s ray ban glasses
* Extracting unstructured data
* Moderation

I put this above the GPT Store because of my confidence (never a good thing to have when making a prediction).

The future is hard to predict but I’m willing to bet this is where we see the most immediate value from

Ok so how do we get started?

The answer is the same as always, go tinker and have fun.

Build an intuition for the tools and find problems you’re passionate about solving.

-------Didn't make the Top 10 cut-------
* GPT 4 Turbo - Multiple Function Calling
* GPT 4 Turbo Knowledge cut off April 2023
* Copywriter shield
* GPT 4 Turbo - (Semi) Deterministic outputs with seed
* GPT 3.5-16K - Fine tuning now available
* ChatGPT - No drop down menu
* GPT 4 Turbo - Output JSON Mode
* Whisper V3 - Soon to the API

[Abdul @_abdul_h]

@ArtsolitNox @levelsio That's a great point! Privacy laws will make it difficult for many, but those that have user agreements that are flexible, or even directly say "your data is mine" will be positioned really well in this race. Exhibit A: https://t.co/z8J061Jfjy

[Abdul @_abdul_h]

@gregisenberg Do you build a community with a potential product in mind? Or perhaps the process of building a community leads you to a product worth building?

[Abdul @_abdul_h]

@businessbarista Take a picture on the equator for memories!

[Abdul @_abdul_h]

@thepatwalls It would be interesting to hear about how he found this business opportunity, and maybe details on the creation process as well? Squeeze in some info on profit margins while we’re at it?

[Abdul @_abdul_h]

@aakashgupta abdul.hersi

[Abdul @_abdul_h]

@gregisenberg Interesting , exact same list I was recommended. Recommendation algorithms suffer from something called “cold start”, when you don’t have enough user data, but I’m sure they could have done better than that. Fun article about cold starts…https://t.co/BuZZdck9My

[Abdul @_abdul_h]

Just like a small adjustment in a golf swing can change the trajectory of a golf ball, even a small advancement in AI can reshape the entire landscape in a hurry. I'll be bookmarking this thread to revisit and compare the good ol' days of AI!

[Abdul @_abdul_h]

GenerativeAI startups' pre-money valuations are up 16%,all other startups are down 24%. 2.6 billion has been invested in 2022,and investment is expected to hit 40 billion by end of year.Expect similar startup success rates-a peek into looming post-mortems.Image via pitchbook(a🧵)

[Abdul @_abdul_h]

starting gun was fired. They are the ones with an expansive database, a solid brand, a broad distribution network. There might be a few new comers that enter the fray, but odds are they get swallowed whole by the incumbents and don’t live to tell the tail...