George Morgan @vr4300 Twitter profile

Pinned Tweet

George Morgan

@vr4300

22 days

1

0

16

Last Seen Profiles

@Mahiratrans_ts

@QualityQueefer

@chefhuyle

@bokeplokalmalam

@galery_basah10

@bokeplokalmalam

@jagoaneonvvip

@ArmaBeyaz

@bokeplokalmalam

@osushivip

@tophung_12

@kuekuenine

@PFFMx

@bokeplokalmalam

@stwmaniax

@reinacrtt

@jandakembangstw

@bokeplokalmalam

@penyukastw21

@DeisiHelemae

@ahmedmona111111

@Omrsevinc

@gobchang91890

@cathslater68

@YingZhao0521

@OussPl

@Kamala_ala

@transscribe

@bokeplokalmalam

@Chitta555

@DeisiHelemae

@Ballslover46747

@SjadAlfsry

@xkfksktlqkfsus

@Keren5816Meral

@pbzpknfwprVUQGS

George Morgan

@vr4300

5 years

@Official64Mate

3

88

354

George Morgan

@vr4300

1 year

@mckaywrigley If you typed even two of those names into ChatGPT I'm sure it could guess the movie by association.

5

0

53

George Morgan

@vr4300

6 years

Sometimes I get Bjarne Stroustrup confused with char *strdup().

2

13

47

George Morgan

@vr4300

5 years

Good progress on the N64 RDP fuzzer this weekend. And the video decoder is working too!

1

8

43

George Morgan

@vr4300

24 days

@theojaffee Still true

0

42

George Morgan

@vr4300

5 years

Here’s a picture of before the layout was cleaned up. You can see the FPGA that’s decoding the RCP’s VI bus and driving a 320x240 LCD panel.

0

5

28

George Morgan

@vr4300

1 year

I think I just discovered a very strange text-davinci-002-render-sha artifact. If you type the letter A (in all caps) 373 times, with a space in between each A, the model will produce a hyper specific answer to a random question. Has anyone seen this before?

6

2

30

George Morgan

@vr4300

2 years

I'd place my bets that discretized (symbolic) systems will have the greatest chance of being the first to demonstrate these properties. Extremely early progress at Symbolica has already demonstrated this. As scale increases, discrete systems may be the most capable models yet.

5

1

23

George Morgan

@vr4300

7 months

Thanks for featuring @symbolica , @FortuneMagazine and @sharongoldman . We're super excited to build the future of AI. A future defined by structure, symbols, logic, and reasoning!

Sharon Goldman

@sharongoldman

7 months

NEW: Vinod Khosla @vkhosla is betting on a former Tesla autopilot engineer @vr4300 , who quit to found @symbolica , which will build small AI models that can reason: “We love people coming from left field,” Khosla said in an exclusive chat with @fortune .

0

6

36

5

25

George Morgan

@vr4300

3 months

@ylecun I think this is approaching the problem from the wrong direction. If you can write a simulator why not just build the model correctly from induction to begin with?

4

0

24

George Morgan

@vr4300

3 months

Great to meet you @mmbronstein . GDL was a huge inspiration to me starting @symbolica . Your kind words mean a lot!

Michael Bronstein

@mmbronstein

3 months

⁦ @symbolica ⁩ is perhaps the most intriguing ML startup out there

2

8

84

0

2

22

George Morgan

@vr4300

1 year

@mineekdev So glad to see people still doing stuff like this. I miss the iPhone Linux days!

0

20

George Morgan

@vr4300

3 months

Making the computer think from our new research lab in London

3

0

20

George Morgan

@vr4300

1 year

@SullyOmarr What is the breakthrough that has enabled you to build an autonomous agent? Surely not using OpenAI APIs right?

2

0

18

George Morgan

@vr4300

11 months

It's becoming increasingly clear to me that @VictorTaelin is onto something huge with HVM. This is going to be a massive deal for AI & computing in the upcoming years.

Taelin

@VictorTaelin

11 months

HVM is becoming the world's fastest λ-calculator! For a perspective, let's perform a Radix Sort on a Scott Tree with millions of ints, vs state-of-art runtimes: JavaScript (V8): 29.081s Haskell (GHC): 11.073s Kind (HVM): 2.514s How is that possible? See below ↓

14

68

629

1

18

George Morgan

@vr4300

6 months

@VictorTaelin Seeding your brain with the correct abstractions to come to powerful conclusions is always the most time consuming part. It's not time wasted!

0

16

George Morgan

@vr4300

3 years

@henrycarless1 @CrabsAndScience @ChemicalKevy @NileRed2 @MrBeast It is called a A100374CT-ND. It's an insulation displacement connector used to attach the leads of the electronic match igniter to the board, as shown in this photo.

0

16

George Morgan

@vr4300

2 months

Proud of have collaborated with Stephen on this work!

Stephen Wolfram

@stephen_wolfram

2 months

What's really going on in machine learning? Just finished a deep dive using (new) minimal models. Seems like ML is basically about fitting together lumps of computational irreducibility ... with important potential implications for science of ML, and future tech...

115

716

4K

1

2

17

George Morgan

@vr4300

7 months

This is precisely why we need to build structured, symbolic models. So that AI can reliably do the things that it's advertised as being capable of.

Arvind Narayanan

@random_walker

7 months

The crappiness of the Humane AI Pin reported here is a great example of the underappreciated capability-reliability distinction in gen AI. If AI could *reliably* do all the things it's *capable* of, it would truly be a sweeping economic transformation.

14

89

460

4

1

15

George Morgan

@vr4300

3 months

@FreyaHolmer Same. They have a great culture, where they trust in their employees to bring only people they trust. It works when it works.

0

15

George Morgan

@vr4300

4 years

@TubeTimeUS Do you ever stop working on projects? The amount of content here is staggering.

2

0

12

George Morgan

@vr4300

4 years

@lexfridman Or, also like computer science, you know you've found the correct tweet, friendship, love, marriage, or meaning of existence when you can reduce the computational complexity to O(1).

2

1

11

George Morgan

@vr4300

4 months

Synthetic data is an incredibly silly concept. If you can build a synthetic data generation engine, by extension it should be possible to directly build a model that understands the innate structure of the task completely skipping the need for more training data to learn it.

Jacques

@JacquesThibs

4 months

I keep saying how underrated it is among alignment researchers

9

6

71

3

0

14

George Morgan

@vr4300

6 years

In heated debate about dongle chaining: "Build me a keyboard to display adapter." - me, confidently "I think that's just a computer." - @harlanhaskins I am foiled again.

0

2

14

George Morgan

@vr4300

2 years

@souplovr23 Plots like these are shadows of beautiful high dimensional structure that we will never be able to visualize. We can pattern match with symbols, but imagine having the ability to see data like this in its true form.

1

14

George Morgan

@vr4300

7 months

Much more to come!

Zanzi Tangle, now at Monoidal Cafe

@tangled_zans

7 months

Lots of category theory in the news lately

0

9

51

1

0

13

George Morgan

@vr4300

3 years

@Rickard0oo @CrabsAndScience @ChemicalKevy @NileRed2 @MrBeast Sure. Here's the code and schematic!

1

0

13

George Morgan

@vr4300

7 months

Amazing how far @symbolica has come since I first met Tim at NeurIPS in 2022. I'm very excited for the upcoming MLST episode on categorical deep learning which will dive into the tech that we are building to pave the way towards structured symbolic cognition in machines!

Machine Learning Street Talk

@MLStreetTalk

7 months

Exciting times ahead for @symbolica - we caught up with Dr. Paul Lessard, a Principal Scientist in their founder team about their neurosymbolic/category theory approach to taming deep learning as a prelude to our special edition we filmed in London. Cameo from @vr4300

0

2

32

0

1

13

George Morgan

@vr4300

2 years

⚠️ Do not make a new programming language. I repeat. Do not make a new programming language. ⚠️

3

0

13

George Morgan

@vr4300

2 years

@0xmaddie_ Or it's just as simple as not having enough infrastructure to handle demand.

0

11

George Morgan

@vr4300

2 months

@bindureddy This is total nonsense, honestly. If they had a reasoning engine, wasting time and energy using it to generate data to train a larger non-reasoning engine model would be an absurd waste.

1

11

George Morgan

@vr4300

11 months

@cstanley Keep going strong. We're proud of you. 💪

2

0

11

George Morgan

@vr4300

3 months

"My CPU is not capable of performing branching or control flow. But don't worry! As long as I run enough instructions it is Turing complete!"

Sasha Rush

@srush_nlp

3 months

Starting out "LLMs just generate the next-word, and therefore cannot..." is a great way to signal you are not making a serious argument.

22

25

455

0

1

10

George Morgan

@vr4300

8 months

@miniapeur Computational statistics.

0

11

George Morgan

@vr4300

18 days

It was inspiring to see much of what I spent 4.5 years of my life working on launched yesterday at the Tesla We, Robot event. I'm proud of my friends and coworkers who pushed it over the finish line. Tesla is building the future. It continues to inspire me to do the same.

0

19

George Morgan

@vr4300

5 months

Yann is now based. I repeat. Yann is now based.

4

0

10

George Morgan

@vr4300

2 years

To reiterate something that I can't stress enough, models that will be capable of dynamic adaptation (true generalization) will necessarily need to define their own loss function. This is incompatible with gradient descent or any gradient based learning models we know of today.

4

1

9

George Morgan

@vr4300

5 years

1D = airline 2D = airplane

0

2

10

George Morgan

@vr4300

1 month

It's time to build machines that are actually capable of reasoning.

1

0

10

George Morgan

@vr4300

2 months

At this point I'm fairly confident that AGI will never be achieved by scaling a single system / architecture. It will most certainly involve bootstraping many systems / architectures with themselves. The first architecture in this process hasn't yet been created. LLMs are not it.

0

10

George Morgan

@vr4300

11 months

@tim_zaman Pretty confident you don't need all the GPUs. You just need a new, powerful technique. :)

0

George Morgan

@vr4300

7 months

Category theory will change machine learning forever. It's the key to being able to describe symbolic reasoning.

0

9

George Morgan

@vr4300

15 days

Solving machine intelligence is the most pressing issue of our time. And the next after that: humanoid robotics

3

0

11

George Morgan

@vr4300

9 months

@svpino It blows my mind that people still think he knows what he's talking about.

4

0

9

George Morgan

@vr4300

1 month

Reasoning cannot magically emerge from data alone. You need the right architecture, the right data, and critically: the ability to interact with a mechanism that validates the result of your reasoning. Can you learn to write programs by looking only at code? No, you need to run

2

0

9

George Morgan

@vr4300

1 year

It gets weirder if you ask it what you just asked it.

2

0

9

George Morgan

@vr4300

1 year

@nice_byte But it also makes incredibly difficult things easy.

1

0

9

George Morgan

@vr4300

3 years

Was a blast building the hardware and firmware for this project. If anyone is interested, I can post a thread on how we pulled it off! Go watch the video!

WilliamOsman

@WilliamOsman

3 years

I finished making all 456 wireless explosive charges for MrBeast's Squid Game. Go watch it!! Ill send a board to a random person who retweets this (does not include explosive)

140

5K

33K

0

1

9

George Morgan

@vr4300

11 months

It's time to accelerate.

0

1

9

George Morgan

@vr4300

1 year

@ID_AA_Carmack It's getting there, especially with embedded support improving. SPARK charges licenses for the compiler, but also provides support. Rust is free, but changes rapidly. The biggest roadblock to using it in safety critical applications is lack of good LLVM support for hardened archs

0

9

George Morgan

@vr4300

1 year

@nice_byte Getting fearless concurrency / async threading working in C / C++ is a complete nightmare, for instance. Not to mention the crazy number of crates Rust has that you can just cargo add and be off to the races instead of dealing with compiler and linker errors for days.

1

0

9

George Morgan

@vr4300

11 months

There won't be *one* way to build AGI, similarly to how there isn't *one* way to derive the value of Pi. Instead, we will see many subtly different algorithms that yield AGI bootstrapping each other until we find an algorithm that optimizes for the hardware we have.

0

1

9

George Morgan

@vr4300

2 months

Could not imagine a worse take tbh

Pliny the Liberator 🐉

@elder_plinius

10 months

AGI will be achieved through prompt engineering at this point. 100% serious.

13

4

57

1

0

10

George Morgan

@vr4300

1 month

I don't buy the "it's the data not the architecture" argument at all. If that were the case, we'd already have AGI.

0

1

9

George Morgan

@vr4300

1 year

@mattmireles @OpenAI @ilyasut @sama What are you talking about? If AGI was achieved internally OpenAI's financial issues would be solved and the board would not have any complaints.

1

0

8

George Morgan

@vr4300

7 months

We're gearing up for the launch of something really amazing. Hope you're all excited for the future of symbolic AI. 👀

1

8

George Morgan

@vr4300

11 months

@hi_tysam Love the premise but this is like saying "the reduction of entropy" in a zip file is a world model. I don't disagree with the premise insofar as this is the fundamental information theoretic container of the world built by the model, but like with zip this isn't actually useful.

4

1

7

George Morgan

@vr4300

9 months

Things like this will be remembered as the most egregious misuse of compute in human history.

JJ

@JosephJacks_

9 months

Holy shit. Meta is training LLaMa 3 on 600,000 H100s.... That's $20 BILLION worth of GPUs.... L.F.G.

79

239

3K

1

0

9

George Morgan

@vr4300

3 years

@chipro Building tools is like building a factory that builds a product, often overlooked entirely or underestimated in complexity. Shoutout to the tool developers! 😁

1

0

George Morgan

@vr4300

5 years

@GregDavill I did this for my board!

1

0

8

George Morgan

@vr4300

7 months

@akbirthko This take is operating at the wrong level of abstraction. Of course the brain is not doing category theory. Think of category theory more like the description language / compiler that provides a definition of what the brain is doing. It's more like the language than the program.

0

8

George Morgan

@vr4300

10 months

The greatest honor one can receive in the field of AI is inventing something that has already been invented by Schmidhuber.

0

8

George Morgan

@vr4300

1 year

@Plinz Almost certainly no chance. Plain and simple: LLMs can't learn online.

1

0

8

George Morgan

@vr4300

5 years

It’s nice and organized now. 😍

2

0

8

George Morgan

@vr4300

25 days

Computer vision didn't work until convolution. Sequence modelling didn't work until recurrence. NLP didn't work until the transformer. But yea, it's just all scale and data right? And also, the transformer is definitely the last architecture.

maharshi

@mrsiipa

25 days

> flow matching outperforms diffusion > llama model outperforms DiTs > scaling data, and compute does this mean data and scaling is the moat, and model architecture do not matter that much?

4

73

0

1

9

George Morgan

@vr4300

5 years

@GregDavill And in dark mode of course. :)

0

7

George Morgan

@vr4300

1 year

Interesting GPT-4V failure.

0

7

George Morgan

@vr4300

2 years

With Yann's latest admittance that LLMs aren't enough to get to AGI, it's surprising to see this kind of attitude. Why are these companies focused on copying what's already being done? A huge focus needs to be placed on figuring out what next-gen architectures will look like.

Sundar Pichai

@sundarpichai

2 years

1/ In 2021, we shared next-gen language + conversation capabilities powered by our Language Model for Dialogue Applications (LaMDA). Coming soon: Bard, a new experimental conversational #GoogleAI service powered by LaMDA.

735

3K

15K

1

0

7

George Morgan

@vr4300

2 months

Victor is one of the only people I know of taking Symbolic AI research seriously. I'd bet on him succeeding. This is the way.

0

7

George Morgan

@vr4300

2 months

My niche symbolic AI twitter audience: what do you think about OpenAI's o1? I think we still have a long long way to go to true reasoning. What do you think?

3

0

7

George Morgan

@vr4300

5 months

I'll be speaking at Open Sauce industry day. If you're there make sure to say hi!

Open Sauce

@OpenSauceLive

5 months

We’ve been cooking up some new and innovative ideas for #OpenSauce2024 and are so excited to announce Industry Day on June 14th!

2

1

86

0

7

George Morgan

@vr4300

21 days

There exists complete and absolute truth. The machines can tell us this absolute truth (a proof) with certainty of its correctness. A proof is a program that we can run. If we build this machine, we unlock the secrets of the computational universe. We must build the orb.

2

0

8

George Morgan

@vr4300

5 years

The basic idea is to bring out the SysAD bus and control signals and then commandeer the bus with the FPGA pictured on the right. This will let the FPGA change writes to the internals of the RCP on the fly. Lets see what happens!

1

0

7

George Morgan

@vr4300

2 years

NeurIPS 2022 was very insightful. It's interesting to see so much attention to "wide" as opposed to "deep" learning models. Hinton's new Forward-Forward Algorithm introduces a possible framework for learning over the forward pass, but still using gradients.

1

0

7

George Morgan

@vr4300

6 months

@jimkxa The hardware is good, if utilized properly. We need better models.

1

0

7

George Morgan

@vr4300

1 year

@KrauseFx I miss this

0

George Morgan

@vr4300

4 months

Chollet describes @symbolica

François Chollet

@fchollet

4 months

Re: the path forward to solve ARC-AGI... If you are generating lots of programs, checking each one with a symbolic checker (e.g. running the actual code of the program and verifying the output), and selecting those that work, you are doing program synthesis (aka "discrete

40

79

857

1

2

7

George Morgan

@vr4300

7 months

@FutureJurvetson We can't continue scaling compute and expecting reasoning to emerge. We must build a formal rigorous study of architecture and then construct models that we know will exhibit properties at scale. In civil engineering you know the bridge you've designed won't collapse at scale.

2

0

7

George Morgan

@vr4300

2 months

@1x_tech Note to self: Symbolica's first humanoid robotics demo won't be this offputting

3

0

7

George Morgan

@vr4300

27 days

This guy cooks

Sam D'Amico

@sdamico

27 days

unboxing

66

7

587

0

7

George Morgan

@vr4300

1 year

@sdand Nobody knows what to do anymore.

0

6

George Morgan

@vr4300

10 months

@brandon_xyzw It's not just that image. It breaks on this one too.

1

0

6

George Morgan

@vr4300

3 years

I feel like I can tick off an engineering "I did that!" box today. My board designs made it to hackaday! 🥲

hackaday

@hackaday

3 years

Engineering on a Deadline for Squid Game

0

6

38

1

6

George Morgan

@vr4300

1 year

The world will change when the day comes that you can load an entire codebase (or multiple) into a language model and make useful inferences from it.

0

6

George Morgan

@vr4300

3 years

Gradient decent

Bart Wronski 🇺🇦🇵🇸

@BartWronsk

3 years

The year is 2021. A flagship product used by almost everyone who has anything to do with images, graphics design, or pixels gets *an option* to draw gradients that are not simply broken.

15

111

703

0

6

George Morgan

@vr4300

1 year

@personofswag Relevant meme for the future.

0

2

6

George Morgan

@vr4300

2 months

@VictorTaelin This is sad but makes total sense. I think you will be very effective putting all your attention back into HVM! It will help HOC move along much faster. And that is all in direct favor of better symbolic AI algorithms too. Really excited to see the Mac Mini cluster come together!

0

6

George Morgan

@vr4300

2 years

I am willing to bet that all of the compute necessary to run a model in real time with emergence indistinguishable from human intelligence is contained in a single RTX 4090 GPU.

2

1

5

George Morgan

@vr4300

4 years

@karpathy Have you tried RTX voice?

0

6

George Morgan

@vr4300

3 years

@ID_AA_Carmack I feel that most good ML papers, and technical papers in general, could be presented in a 15 minute 3Blue1Brown style video that make the actual-thing-that-helps extremely obvious. If the paper can’t be presented in such a format, it is likely that it is just fluff.

0

6

George Morgan

@vr4300

7 years

I'm starting to write an N64 emulator in Rust! :)

0

1

6

George Morgan

@vr4300

3 years

@ID_AA_Carmack I think there should be an "ImageNet in the smallest possible self contained executable" contest. Zip everything up into a single CPU only executable and see who can achieve the best runtime vs executable size ratio.

0

6

George Morgan

@vr4300

1 year

@pmddomingos Synthetic data makes less than 0 sense. It indicates a huge flaw with the foundations of the models.

3

0

6

George Morgan

@vr4300

5 months

Congrats Masha! Really excited to be working with you and the team at Day One. And equally excited for the future of your fund. 💪🎉

Masha Bucher

@mashadrokova

5 months

Today, I’m proud to announce @DayOneVC ’s $150M fund III that brings our AUM to over $450M. Our north star remains the same: we will keep betting on the most exceptional founders of our time working on the biggest ideas possible.

86

39

566

0

3

6

George Morgan

@vr4300

1 year

@levie AI is the ultimate example of this.

0

George Morgan

@vr4300

3 years

Go watch @WilliamOsman 's documentary on how we saved @MrBeast 's Squid Game!

0

1

6

George Morgan

@vr4300

2 months

Incredible post by @sedielem that showcases why visualizing data using different frames of reference is so powerful!

Sander Dieleman

@sedielem

2 months

Diffusion is the rising tide that eventually submerges all frequencies, high and low 🌊 Diffusion is the gradual decomposition into feature scales, fine and coarse 🗼 Diffusion is just spectral autoregression 🤷🌈

33

164

1K

0

7

George Morgan

@vr4300

11 months

History repeats itself. Humans are so easily fooled.

François Chollet

@fchollet

11 months

I'm old to enough to remember when *GPT-2* was so dangerous it couldn't be released.

8

35

452

0

6

George Morgan

@vr4300

4 years

Cool new translucent SP, with custom battery cell, IPS display, and translucent atomic purple shell.

3

0

6

George Morgan

@vr4300

1 year

@killroy42 @mckaywrigley @Ciaran2493 Agreed. It's certainly impressive that it can extract the spatial information from the arrows but its understanding of the hierarchy of the story almost certainly comes from the information about the movie in its text training set.

1

0

5