Cedric Knapp

Computer Scientist, Virtual Reality Developer and Enthusiast, AI inquisitive, Indie Game Developer

[Cedric Knapp @CedricKnapp]

As promised, the first beta of my VR project: "Single Shot Smash!" is up on Side Quest! Currently waiting on the approval process👀
#VR #IndieGameDev #unity3d

[Cedric Knapp @CedricKnapp]

So excited to have my hands on the Quest 3, the MR capabilities are going to be fun to experiment with #indiedev

[Cedric Knapp @CedricKnapp]

A stance in solidarity against Unity's recent predatory and hurtful monetization changes

#Unity
https://t.co/AuREwxJhRP

[Cedric Knapp @CedricKnapp]

A sad day for unity developers... I have a hunch unreal & godot will be picked up more due to this... this is another example of greed (due to unity going public) causing a company to make BAD decisions for the long term...
#gamedevelopment #indiedev

[Kenney @KenneyNL]

Unity will soon charge devs per-install after a certain threshold. I'd like to express my concerns regarding privacy (tracking installs/downloads), preservation and changing the whole business model while devs are using the engine for their projects. 🫤

https://t.co/vic6cei5Vm

[Cedric Knapp @CedricKnapp]

Playing around with creating some sharers in #unity3d with unity muse #AI

[Cedric Knapp @CedricKnapp]

@evsevsevsevsevs This is too real

[Cedric Knapp @CedricKnapp]

This just gave me such nostalgic feelings

[Neal Agarwal @nealagarwal]

Added Club Penguin rooms to the museum 🥲

[Cedric Knapp @CedricKnapp]

I just got closed beta access to muse chat, Unity's ChatGPT like inference trained on its internal documentation. Super excited to see how this begins to lower the barrier to learning game development for many people!

#unity3d #ai #gamedev

[Cedric Knapp @CedricKnapp]

I was just granted closed beta access to Unity's new neural network inference library, Sentis. Excited to poke around and experiment with it! #unity3d #ai #gamedev

[Cedric Knapp @CedricKnapp]

OpenAI's 2023 roadmap was released but promptly removed. Here is the archived version if you're still interested!

Original link:
https://t.co/CvX6BctdZr

Archived Link:
https://t.co/aE5XT0WWSf

#AI #OPENAI

[Cedric Knapp @CedricKnapp]

@EddCoates @forkball_vg Very cool! Happy to see the VR section too!

[Cedric Knapp @CedricKnapp]

This FSR/DLSS alternative for #VR for XR2 powered devices allows for much higher visual fidelity to be streamed to the device with a potentially lower latency. Great for increased quality at a lower latency (reducing potential motion sickness)

[Guy Godin @VRDesktop]

Today we are releasing a really cool update that adds upscaling to your streamed PCVR games in Virtual Desktop. We've worked with Qualcomm to bring something similar to FSR but runs on the XR2.

It is called Snapdragon Game Super Resolution (SGSR) and it can work along with SSW

[Peter Nixey @peternixey]

If GPT4 is designed to give you the “next most probable word” then why don’t its answers feel average?

They feel well above average. Which is odd.

Most of us understood GPT4 to be a “word frequency machine”. One that ingested billions of words from the internet and learned how to replay them intelligently. A “stochastic parrot” giving us the most probable, most average, responses to our questions.

But the more you use it, the harder it gets hard to avoid the fact that it's not average. It’s top decile. And it seems to understand much more than mere word order.

GPT3 first went viral creating poems and lyrics to songs. Which could, maybe, have been the result of a stochastic word-order machine. So let's try its successor on a harder lyrical challenge.

🧑🏻‍🦰GPT4, In the classic song, American Pie, one verse refers to dancing in a gym. Assume the dance follows a hostage situation where the police shot 1,000 rounds through the gym’s windows. Which lyric would most need to change?

🤖 The lyric to change is: "You both kicked off your shoes". Given the context of shattered glass from police shooting through the windows, it would be unlikely and dangerous for the couple to be dancing without their shoes on.

You'd be hard pushed to find any text on the internet that answers my weirdo question. It can’t just be “autocompleted”.

And answering it requires a lot of understanding:

- the consequences of a shootout
- what each lyric means
- how all the consequences affect all the lyrics
- which lyrics are most affected
- how bare feet are vulnerable to broken glass

Show me a parrot that can do this and I’ll show you a big-old bag of circus money.

The evidence that it knows more than just word order isn’t just anecdotal. If you do a little maths you realise that doing an exhaustive word-frequency-analysis also requires an impossibly large dataset.

There are ~500,000 words in the English language so a sentence of just 15 words has more permutations than there are atoms in the universe. And GPT4 can handle instructions of up to 8,000 words.

When you bring these ideas together, they hint at an astonishingly powerful conclusion:

“When we train an LLM to accurately predict the next word, what we are doing is learning a world model. It may look like we’re learning statistical correlations in text but it turns out that what's actually learned is some representation of the process that produced the text. The text is a projection of the world model. The neural network is not learning the sequence of text, it's learning the model of the world that is projected down into that text”.
- Ilya Sutskever (@ilyasut) , Cofounder of @OpenAI

So it is indeed predicting the next most probable word. But it's predicting it based not on a model of our language, but on a model of the world our language describes. Which is far, far more powerful.

(I write on AI from a technical and product perspective. If you find that interesting please do follow me for more)

[Cedric Knapp @CedricKnapp]

Wow today is a happy day, backup codes are now account synced in google authenticator. Only took 13 years 😅😬🥴
https://t.co/yiRfgi7uy5

[Cedric Knapp @CedricKnapp]

GptCache - a very cool open-source project aimed at utilizing a semantic caching approach for reducing latency and token costs associated with interacting with various LLM (chatGPT) API calls.
https://t.co/lqWkbjLixz

[Peter Nixey @peternixey]

I'm in the top 2% of users on StackOverflow. My content there has been viewed by over 1.7M people. And it's unlikely I'll ever write anything there again.

Which may be a much bigger problem than it seems. Because it may be the canary in the mine of our collective knowledge.

A canary that signals a change in the airflow of knowledge: from human-human via machine, to human-machine only. Don’t pass human, don’t collect 200 virtual internet points along the way.

StackOverflow is *the* repository for programming Q&A. It has 100M users & saves man-years of time & wig-factories-worth of grey hair every single day.

It is driven by people like me who ask questions that other developers answer. Or vice-versa. Over 10 years I've asked 217 questions & answered 77. Those questions have been read by millions of developers & had tens of millions of views.

But since GPT4 it looks less & less likely any of that will happen; at least for me. Which will be bad for StackOverflow. But if I'm representative of other knowledge-workers then it presents a larger & more alarming problem for us as humans.

What happens when we stop pooling our knowledge with each other & instead pour it straight into The Machine? Where will our libraries be? How can we avoid total dependency on The Machine? What content do we even feed the next version of The Machine to train on?

When it comes time to train GPTx it risks drinking from a dry riverbed. Because programmers won't be asking many questions on StackOverflow. GPT4 will have answered them in private. So while GPT4 was trained on all of the questions asked before 2021 what will GPT6 train on?

This raises a more profound question. If this pattern replicates elsewhere & the direction of our collective knowledge alters from outward to humanity to inward into the machine then we are dependent on it in a way that supercedes all of our prior machine-dependencies.

Whether or not it "wants" to take over, the change in the nature of where information goes will mean that it takes over by default.

Like a fast-growing Covid variant, AI will become the dominant source of knowledge simply by virtue of growth. If we take the example of StackOverflow, that pool of human knowledge that used to belong to us - may be reduced down to a mere weighting inside the transformer.

Or, perhaps even more alarmingly, if we trust that the current GPT doesn't learn from its inputs, it may be lost altogether. Because if it doesn't remember what we talk about & we don't share it then where does the knowledge even go?

We already have an irreversible dependency on machines to store our knowledge. But at least we control it. We can extract it, duplicate it, go & store it in a vault in the Arctic (as Github has done).

So what happens next? I don't know, I only have questions.

None of which you'll find on StackOverflow.

(I write on AI from a technical and product perspective. If you find that interesting then please do follow me for more)

[𝙏𝙝𝙧𝙞𝙡𝙡 @Thrilluwu]

I've been witnessing a lot of transphobia on the internet lately.

I'll say this once: Freedom of expression will always prevail. You have a choice; give people freedom or try and take it away.

I am an ally and I always will be. Let freedom of expression prevail.

[Cedric Knapp @CedricKnapp]

@evsevsevsevsevs It's crazy to me this isn't a staple food for people

[Cedric Knapp @CedricKnapp]

GPT-4 powered Copilot just released and looks really cool. More integrated in VScode & Visual Studio IDE's. #AI #development
https://t.co/VlrDF401hW

[Leonardo.Ai @LeonardoAi]

Mesh-aware texture generation will fire up the growth of #UGC and accelerate 3D production.

Look out for our Blender plugin coming later this month, alongside our front-end platform implementation, as well as API access. 🚀🚀🚀

Early access
https://t.co/xVWUKnZmwe

Model + text prompt >> UV textures, normals, roughness, displacement.