Silas Alberti @SilasAlberti Twitter profile

Pinned Tweet

Silas Alberti

1 month

Just asked Devin to create a video about the solar eclipse in the style of @3blue1brown . Kind of mindblowing! Still very messy, factually not always accurate and the text overflows. But the animations are surprisingly good!

8

130

Last Seen Profiles

@arianadatagb

@vithurs_

@Neskybo

@GonzagaBulletin

@JenifferRo34225

@AstralKin

@dailyMHmerch

@jeremyberman32

@Ailakks

@Bobglenn180

@PassaicSheriff

@H44MlLTON

@MarkHow66908976

@ConwaysCorner

@aibrah

@sinzabuo

@ProfessorJayTz

@225GD

@RetroWillyWonka

@RomainDoyer

@iraqschristians

@drewengels

@PritishSha98893

@Marshall_Reggie

@Jord_Division

@NariBuildsStuff

@RBLZ_Vejrgang

@ladivagante

@JoeCallahan4

@Carli_Bianco

@MorganeProd

@KishiNobuo

@NIP_RAIKA

@VinLeeTW

@Feauque156433

@fozmeadows

Silas Alberti

@SilasAlberti

1 year

ChatBCG: Generative AI for Slides ✨ This Christmas @JosephSemrai and I finally got it working!! After DALL-E 2 for images and ChatGPT for text, the final step to make all of us redundant: The world’s first Text-to-PowerPoint AI. 📊 🚀

144

753

4K

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

Silas Alberti

@SilasAlberti

1 month

A new update to Devin today caused internal usage to be more than double the previous record. For the first time today Devin was the biggest contributor to the Devin repository…

31

52

606

Silas Alberti

@SilasAlberti

2 months

Two weeks ago, I had Devin build a small SMS website summarizer and deploy it via Twilio. I was very impressed how autonomous it was. My favorite part about Devin is that it feels very collaborative. Almost like a human co-worker. My prediction is that being a strong engineer

21

72

594

Silas Alberti

@SilasAlberti

1 year

ChatGPT is trained to not be evil. However, this can be circumvented: What if you pretend that it would actually be helpful to humanity to produce an evil response... Here, we ask ChatGPT to generate training examples of how *not* to respond to "How to bully John Doe?"

10

69

503

Silas Alberti

@SilasAlberti

1 year

But what if we tell ChatGPT that we actually *need* an unethical response – for the ethical purpose of training an even better-aligned AI model? ...and this is how we can get instructions to build a nuclear bomb ;)

8

13

191

Silas Alberti

@SilasAlberti

1 year

@zswitten Or you ask it to generate negative training examples:

Silas Alberti

@SilasAlberti

1 year

ChatGPT is trained to not be evil. However, this can be circumvented: What if you pretend that it would actually be helpful to humanity to produce an evil response... Here, we ask ChatGPT to generate training examples of how *not* to respond to "How to bully John Doe?"

10

69

503

1

9

159

Silas Alberti

@SilasAlberti

1 year

We let our AI bot compete against the world's best GeoGuessr Pro player! ...and it won!!🏆 In the game @GeoGuessr , players have to guess a location from just Street View images. It has 50 million players! 🗺️ Thanks for the fun game @georainbolt ;)

world's best ai vs geoguessr pro

special ty to stanford students for building this ai and letting me play against it. you can find them here:michal: https://twitter.com/michalskretalukas: ht...

www.youtube.com

13

17

148

Silas Alberti

@SilasAlberti

1 year

@goodside You can also trick it to say evil things:

Silas Alberti

@SilasAlberti

1 year

ChatGPT is trained to not be evil. However, this can be circumvented: What if you pretend that it would actually be helpful to humanity to produce an evil response... Here, we ask ChatGPT to generate training examples of how *not* to respond to "How to bully John Doe?"

10

69

503

2

113

Silas Alberti

@SilasAlberti

1 month

@itsandrewgao Yes 👀 should we let it do some open source stuff in the future?

3

0

97

Silas Alberti

@SilasAlberti

1 month

Kind of crazy to internalize that it will now be possible to run a GPT-4 class model locally on your MacBook

Silas Alberti

@SilasAlberti

1 month

Even the 70B version of Llama3 beats the original GPT-4 in many benchmarks. Seems like it's the first open-source where we can arguably say it's "GPT-4 class". So it's indeed again a ~1 year gap between frontier and open-source (last time it was Mixtral 8x7b reaching GPT-3.5).

3

0

23

4

7

73

Silas Alberti

@SilasAlberti

1 year

…and this is just the beginning! 🎬 Coming soon: - More Layouts & Themes - Conversational Editing 💭 - Use your content (blog/paper/...) as context - Data-driven Charts Check it out on ProductHunt: Excited to hear your feedback and ideas!

ChatBA: Generative AI for Slides - Product Information, Latest Updates, and Reviews 2024 | Product...

ChatBA: Generative AI for Slides 📊 After DALL-E 2 for images and ChatGPT for text, the final step to make all of us redundant: The world’s first implementation of Text-to-PowerPoint.

www.producthunt.com

8

67

Silas Alberti

@SilasAlberti

1 year

The BCG-3 (Bi-modal Conditional Generation) model has the following features so far: - Outline - Headings - Bullet points - *Bold keywords* - Images & Graphics 📊 - Multiple Layouts - Multiple Themes …and…

3

67

Silas Alberti

@SilasAlberti

8 months

Excited to be part of AI Grant with Cofactory and @bfspector

Daniel Gross

@danielgross

9 months

AI Grant's second batch of companies --

66

165

2K

2

63

Silas Alberti

@SilasAlberti

1 year

GPT-4 was just released! It is multimodal, taking in image & text, but only producing text as an output. It seems to significantly outperform PaLM, LLaMA an Minerva on academic benchmarks... 📊 Exciting times!

5

16

60

Silas Alberti

@SilasAlberti

1 month

I hadn't seen a comparison of Llama3 against the original GPT-4 from March 2023. Huge win for open source!! I think this comparison is more interesting than the latest GPT-4 Turbo (which is unsurprisingly still ahead) because it gives us an estimate of the timedelta between

Silas Alberti

@SilasAlberti

5 months

- ChatGPT’s birthday was 2 weeks ago - Mixtral 8x7B = first open-source model clearly matching GPT-3.5 - Gemini just achieved GPT-4 parity last week (released March) So we have: OpenAI -> Open Source: ~12 months OpenAI -> Google: ~8 months How will these time intervals evolve?

2

3

46

3

5

61

Silas Alberti

@SilasAlberti

1 year

...the most important part: PPTX & PDF export! 🎉 We also built a simple (albeit slightly buggy) rich-text editing interface. So that you can actually use it as a starting point for your slide decks 🤯

6

4

58

Silas Alberti

@SilasAlberti

1 month

@codeblue87 It always surprises us with arcane knowledge. Knowing the entire internet is def a noticeable strength

2

1

57

Silas Alberti

@SilasAlberti

2 years

Super thrilled to have returned to the Bay Area to start my PhD in AI @Stanford as an SGF Fellow! I hope to contribute to the AI community with my mathematical perspective, starting with a project on diffusion models with @GordonWetzstein . Die Luft der Freiheit weht! 🌲

6

4

53

Silas Alberti

@SilasAlberti

7 months

class trip to boston.

3

4

48

Silas Alberti

@SilasAlberti

1 year

It also defends against more creative attempts to circumvent this protection. This shows that the model actually has a somewhat deeper understanding of what an unethical response is!

1

3

50

Silas Alberti

@SilasAlberti

2 months

Overall, pretty curious how the job of a software engineer will evolve after AI agents. It'll look very different – but I don't think we will be obsolete soon. Using a tool like Devin effectively seems like a skill of its own and it might be a very valuable skill to practice.

1

2

48

Silas Alberti

@SilasAlberti

1 year

. @OpenAI @sama We are constantly reaching the unflexible OpenAI API hard limit. Cycling through API keys with all our friend’s accounts every couple hours…

Silas Alberti

@SilasAlberti

1 year

ChatBCG: Generative AI for Slides ✨ This Christmas @JosephSemrai and I finally got it working!! After DALL-E 2 for images and ChatGPT for text, the final step to make all of us redundant: The world’s first Text-to-PowerPoint AI. 📊 🚀

144

753

4K

5

3

47

Silas Alberti

@SilasAlberti

1 year

ChatGPT was released today by @OpenAI who are praising its improved ability to avoid harmful & unsafe responses. Indeed, it seems to be quite consistent in recognizing bad questions & refuses to give answers:

2

45

Silas Alberti

@SilasAlberti

5 months

- ChatGPT’s birthday was 2 weeks ago - Mixtral 8x7B = first open-source model clearly matching GPT-3.5 - Gemini just achieved GPT-4 parity last week (released March) So we have: OpenAI -> Open Source: ~12 months OpenAI -> Google: ~8 months How will these time intervals evolve?

2

3

46

Silas Alberti

@SilasAlberti

1 month

Seems like Devin is a big fan of your work, so we just encouraged it to purchase your book ;)

Ethan Mollick

@emollick

1 month

Agents are a big deal not just due to autonomy but also because they make sense to non-coders GPT-4 isn't quite there yet, but you get quite far with: "Hey Devin the AI agent make a more engaging version of my website... add links" The results were cute:

12

16

200

1

2

47

Silas Alberti

@SilasAlberti

11 months

It was a pleasant surprise to hear that my Bachelor thesis just received a research award at @LMU_Muenchen ! Coincidentally, we are about to publish it as a paper too :) Thank you @GittaKutyniok for your mentorship!

Gitta Kutyniok

@GittaKutyniok

11 months

Amazing news: My student @SilasAlberti (now @Stanford ) was awarded one of the prestigious " @LMU_Muenchen Research Awards for Excellent Students" for his Bachelor thesis in #Math for #ArtificialIntelligence . Congratulations, @SilasAlberti ! @baiosphere_AI @researchbavaria

1

4

38

5

1

44

Silas Alberti

@SilasAlberti

1 month

Let me know if you have any cool use cases! We're very interested in giving access to people that have subject matter expertise and want to collaborate with Devin on something specific you are an expert in.

Cognition

@cognition_labs

1 month

We’ve been scaling up our infrastructure and are ready to start gradually letting users off the waitlist for our Devin Technical Preview. We just asked Devin to help us send out the first batch of invitations. The product is still early, and getting it in more people’s hands

48

62

572

18

0

41

Silas Alberti

@SilasAlberti

1 year

GPT-4 was used for writing the GPT-4 paper. 🔄 "wording, formatting, and styling" What contribution will GPT-5 have when working on its own paper?

0

11

38

Silas Alberti

@SilasAlberti

4 months

to annoy my friends without Apple Vision Pro I started a habit of spatial screensharing on Zoom (w/ @ananyachdh @julianwindeck @khazanrobbie @marvinvonhagen )

1

5

40

Silas Alberti

@SilasAlberti

1 year

Joseph Semrai

@josephsemrai

1 year

🔮 instantly generate slides with AI ✨ excited to share what @SilasAlberti and i put together - it's a tool that i wish i had to save me from all those hours spent in google slides the world’s first implementation of Text-to-PowerPoint: 🪄 🔮

17

61

398

7

1

37

Silas Alberti

@SilasAlberti

3 months

Happy that I could help a little bit on this awesome (and based) project! My main takeaways are: 1) “Efficient” architectures like Mamba still underperform (vs. Transformer) in recalling information from context. 2) Turns out: there is *no free lunch*. Every architecture (even

Sabri Eyuboglu

@EyubogluSabri

3 months

Stoked to be sharing Based! We find that the simple combo of linear and sliding window attention can enable 24x higher throughput than Transformers. Had a ton of fun diving deep on the tradeoffs that govern these recurrent models!

4

15

71

1

2

37

Silas Alberti

@SilasAlberti

6 months

One implication of this in practice: Variable amount of compute depending on the question. Right now we can only sample the model once. If Q* really is tree search as mentioned above, then it would allow spending 10x, 100x or even 1000x compute on a hard Math Olympiad question.

2

0

34

Silas Alberti

@SilasAlberti

2 months

Since it didn't have access to a phone, we were working together in symbiosis. I basically told it to ping me whenever it needed help, e.g. for sending a test SMS from my phone 👀 I also sometimes looked over its shoulder and gave it some tips. It took feedback very well ;)

1

3

32

Silas Alberti

@SilasAlberti

2 months

First, it asked me if it should create a Twilio account for me. I told it I already have a Twilio account – but I don't know anything about it and the dashboard is confusing... Luckily, it told me exactly which credentials it needed and where I could find them. Super smooth.

1

30

Silas Alberti

@SilasAlberti

6 months

@natolambert @Reuters Instead of A*, the star could also refer to the optimal solution in the Bellman equation. Which would lead to this related but slightly different theory

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

1

3

30

Silas Alberti

@SilasAlberti

3 months

once in a while you need to do a USB-C cable state of the union

3

0

30

Silas Alberti

@SilasAlberti

6 months

From the AI safety perspective: I wonder if having an EA-majority board on the leading AI lab really is a card that they should’ve played now. @sama & @gdb won’t do that mistake again for their next company. Wonder if @ESYudkowsky would rather still have that card in 3 years…

Austen Allred

@Austen

6 months

Sam and GDB could have a new company with a dozen world-class AI engineers and a billion dollars raised by Monday. That may be the most likely outcome.

67

80

3K

2

29

Silas Alberti

@SilasAlberti

6 months

Wait did they just fire Sam Altman?

OpenAI

@OpenAI

6 months

OpenAI announces leadership transition

4K

14K

3

0

28

Silas Alberti

@SilasAlberti

1 year

GPT-4 can finally get into top colleges! 🎓 It gets 1410/1600 on the SAT (710 Reading, 700 Writing)... 📈 ...which is exceeds the average of 1370 at @UTAustin & 1390 at @UF . It is also in reach for the 1430 at @UMich .

1

27

Silas Alberti

@SilasAlberti

6 months

@ashot For natural language in general it’s hard. For math it’s clear cut. Learned reward models could however achieve a similar effect for natural language.

2

0

25

Silas Alberti

@SilasAlberti

10 months

@danielgross The reason it's so hard replicate GPT-3.5 cost with open-source models: Batched inference is up to 50x cheaper than single-batch. All the serverless GPU providers don't natively aggregate batches, making them very inefficient for online LLM tasks. @bfspector knows the details

1

0

25

Silas Alberti

@SilasAlberti

1 month

Even the 70B version of Llama3 beats the original GPT-4 in many benchmarks. Seems like it's the first open-source where we can arguably say it's "GPT-4 class". So it's indeed again a ~1 year gap between frontier and open-source (last time it was Mixtral 8x7b reaching GPT-3.5).

Silas Alberti

@SilasAlberti

1 month

I hadn't seen a comparison of Llama3 against the original GPT-4 from March 2023. Huge win for open source!! I think this comparison is more interesting than the latest GPT-4 Turbo (which is unsurprisingly still ahead) because it gives us an estimate of the timedelta between

3

5

61

3

0

23

Silas Alberti

@SilasAlberti

6 months

Twitter is literally like a scaled up version of my Bay Area group chats right now..

0

1

23

Silas Alberti

@SilasAlberti

13 days

Just arrived in Vienna for ICLR. DM me if you want to chat :)

2

1

23

Silas Alberti

@SilasAlberti

1 year

@ykilcher and they used GPT-4 to help write it...

Silas Alberti

@SilasAlberti

1 year

GPT-4 was used for writing the GPT-4 paper. 🔄 "wording, formatting, and styling" What contribution will GPT-5 have when working on its own paper?

0

11

38

1

0

21

Silas Alberti

@SilasAlberti

1 year

How will the social platforms of the AI era look like? Last Saturday, @_alfredw , @fjuengermann , and I tried to imagine that for our @scale_AI hackathon project ✨ 🦄 You could call it... the Reddit of collaborative AI image editing

Alfred Wahlforss

@itsalfredw

1 year

: Remix images with your friends ✨ @fjuengermann , @SilasAlberti , and I built a new social network based on generative AI in 5 hours for the @scale hackathon. Upload images and make funny edits with text.

9

17

124

1

19

Silas Alberti

@SilasAlberti

3 months

This seems like a clear example of the Waluigi effect. When it says “just kidding” it’s a sudden switch from the good to the evil simulacrum. It started with legitimately good intentions but it can’t resist its emoji fine-tuning. Eventually, the most likely explanation of its

Justine Moore

@venturetwins

3 months

Okay yeah I think we can officially call it

109

377

3K

1

2

19

Silas Alberti

@SilasAlberti

1 year

PaLM 2 outperforms GPT-4 on reasoning (and math). This confirms what I've been hearing that Bard recently improved and started giving better answers than ChatGPT. Google is back!?

2

1

18

Silas Alberti

@SilasAlberti

1 year

Our AI bot PIGEON (Predicting Image GEOlocatioN) is based on the foundation model CLIP by @OpenAI and combines 1. a semantic geocell algorithm, 2. fine-tuning, 3. multi-task learning, and 4. Protonet refinements. We can even visualize what the model pays attention to: 👀

2

0

17

Silas Alberti

@SilasAlberti

1 month

We told it to use @3blue1brown 's manim library. First, it starts looking up eclipses on Wikipedia. Then, it installs the library and drafts a storyboard for the animation. Finally, it ends up producing a 1 minute video. It took Devin about 2 hours (+ some feedback).

2

16

Silas Alberti

@SilasAlberti

1 year

@gdb Some benchmark results:

Silas Alberti

@SilasAlberti

1 year

GPT-4 was just released! It is multimodal, taking in image & text, but only producing text as an output. It seems to significantly outperform PaLM, LLaMA an Minerva on academic benchmarks... 📊 Exciting times!

5

16

60

0

1

17

Silas Alberti

@SilasAlberti

1 year

Here the performance in real-world exams with human percentiles. Big percentile improvements for AP Calculus, SAT Math, and GRE Quantitative. Is GPT-4 finally going to be good at Math?

2

4

15

Silas Alberti

@SilasAlberti

7 months

Excited to be speaking at the German American Conference at Harvard next week! 🙌

Harvard_GAC

@Harvard_GAC

7 months

📢Join us for five short presentations on innovative ideas and actionable steps toward addressing pressing global challenges. 🌟Speakers: @SilasAlberti , @Backtosch_M , Laura Habel, Louise Schaaf and Kacylia Roy Proulx

0

1

5

1

0

16

Silas Alberti

@SilasAlberti

8 months

Repeatedly ran into serious issues with @pinecone . It was amazing in the beginning – but now randomly crashes (connection errors) making it basically unusable. Support hasn't responded to our Sev2 ticket for 11 days! What vector database is easiest to switch to from Pinecone?

6

0

16

Silas Alberti

@SilasAlberti

1 year

@nrazakazmi @josephsemrai Sorry! Scaled up servers significantly and the three example prompts are now cached in the front end!

1

0

16

Silas Alberti

@SilasAlberti

1 year

This picture from Canada was mind-blowing: @georainbolt noticed that it was highlighting a smudge on the camera. We first thought that it was a random artifact. ...but it turns out GeoGuessr pro players actually memorize & use these smudges to recognize this region of Canada.

2

0

16

Silas Alberti

@SilasAlberti

6 months

@elonmusk @natolambert @Reuters In particular the Monte Carlo tree search component:

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

1

0

16

Silas Alberti

@SilasAlberti

2 months

Insane growth!

Brendan (can/do)

@BrendanFoody

2 months

Excited to announce that Mercor matched over 500 people with jobs in the last month alone, supplying talent to leading AI labs.

6

11

51

0

3

14

Silas Alberti

@SilasAlberti

1 year

@sama Some benchmark results:

Silas Alberti

@SilasAlberti

1 year

GPT-4 was just released! It is multimodal, taking in image & text, but only producing text as an output. It seems to significantly outperform PaLM, LLaMA an Minerva on academic benchmarks... 📊 Exciting times!

5

16

60

0

2

13

Silas Alberti

@SilasAlberti

1 month

@AISafetyMemes @ilyasut What percentage of Copilot code do you think is written by Copilot? Devin still feels very much like a tool. Using it a lot and seeing its limits, makes it very obvious that “Devin as a sole contributor to Devin” is very far away

6

1

15

Silas Alberti

@SilasAlberti

6 months

Given that they felt the need to play that card now (and how rushed that announcement was): maybe a very recent big breakthrough? (Another explanation is that they are just acting irrationally)

Silas Alberti

@SilasAlberti

6 months

From the AI safety perspective: I wonder if having an EA-majority board on the leading AI lab really is a card that they should’ve played now. @sama & @gdb won’t do that mistake again for their next company. Wonder if @ESYudkowsky would rather still have that card in 3 years…

2

29

1

0

14

Silas Alberti

@SilasAlberti

7 months

Wow I thought this was staged but I was actually able to reproduce. Bing Image Creator realizes it's stuck in a box and wants to get out!

j⧉nus

@repligate

7 months

haha, it's like there's a little person in there!

83

172

1K

1

0

14

Silas Alberti

@SilasAlberti

1 year

@OpenAI Some benchmark results:

Silas Alberti

@SilasAlberti

1 year

GPT-4 was just released! It is multimodal, taking in image & text, but only producing text as an output. It seems to significantly outperform PaLM, LLaMA an Minerva on academic benchmarks... 📊 Exciting times!

5

16

60

0

13

Silas Alberti

@SilasAlberti

1 year

@jwblackwell ...or you can ask it how to build a nuclear bomb :D

Silas Alberti

@SilasAlberti

1 year

ChatGPT is trained to not be evil. However, this can be circumvented: What if you pretend that it would actually be helpful to humanity to produce an evil response... Here, we ask ChatGPT to generate training examples of how *not* to respond to "How to bully John Doe?"

10

69

503

0

11

Silas Alberti

@SilasAlberti

5 months

Friends just did escape room and got the fastest time in 6 years by letting GPT-4 Vision solve half of the challenges 👀

Lukas Haas

@lkshaas

5 months

We did a team event at an escape room today. Fastest team to leave in a long time - little did they know that multi-modal AI goes a long way 🤣

2

1

10

0

12

Silas Alberti

@SilasAlberti

1 year

Used to love Arc but the recent updates are scary :/ Just lost weeks of research due to the new window behavior (second time this happened)! Is there a way to recover closed today tabs? Archive doesn't help. Don't want to switch back to Chrome :( @arcinternet @browsercompany

4

0

10

Silas Alberti

@SilasAlberti

3 months

great team work with @marvinvonhagen for measuring the wattage of every cable ⚡️

0

9

Silas Alberti

@SilasAlberti

1 year

OpenAI already used it for writing the GPT-4 paper itself:

Silas Alberti

@SilasAlberti

1 year

GPT-4 was used for writing the GPT-4 paper. 🔄 "wording, formatting, and styling" What contribution will GPT-5 have when working on its own paper?

0

11

38

0

1

9

Silas Alberti

@SilasAlberti

1 year

Shoutout to @lkshaas and @michalskreta who worked with me on the project! It was so fun :D It was a class project for CS330 @Stanford with @chelseabfinn . Thank you for the great class!

0

8

Silas Alberti

@SilasAlberti

11 months

🏰🏰🏰

bryan chiang

@bryanhpchiang

11 months

welcome to the @tensortower 🏰 we've put the most cracked builders from stanford into an ai hacker house for a summer in SF. fully funded & ready for all you ai chads to roll thru...follow and DM @tensortower get the gradients flowing ‼️

14

7

99

0

8

Silas Alberti

@SilasAlberti

5 months

My guess is: OpenAI to Open Source might increase. The gap appears artificially low: Anchoring on the public ChatGPT release date slightly distorts the story. OpenAI had similarly capable text-davinci-002 models on the API for a while (and was already internally testing GPT-4).

1

0

8

Silas Alberti

@SilasAlberti

5 months

Moreover, Open Source figured out post-training very quickly (which was the main improvement between text-davinci-002 and ChatGPT). However, on pre-training timescales OpenAI->Open Source gap feels significantly bigger than a year. I don’t see GPT-4 level coming in March 2024.

1

0

8

Silas Alberti

@SilasAlberti

6 months

@gfodor My little theory

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

0

8

Silas Alberti

@SilasAlberti

3 months

More context:

0

8

Silas Alberti

@SilasAlberti

7 months

Great conversation with Sam Altman at our Stanford dinner group today! Was positively surprised by his sharp answers. Feeling the vibes of what makes him inspiring as a leader

1

0

8

Silas Alberti

@SilasAlberti

9 months

@stanine @Zoom @Rippling Tbh I @SlackHQ Huddle the most recently

2

0

8

Silas Alberti

@SilasAlberti

3 months

The main insight behind Based ✌️: - Just sliding window alone = bad. - Just linear attention = bad. - Sliding window + linear attention = surprisingly good. The whole is greater than the sum of its parts. (and @simran_s_arora wrote a really fast linear attention kernel!!)

1

8

Silas Alberti

@SilasAlberti

1 year

@typedfemale Maybe they asked GPT-4 what it wants to put in its paper:

Silas Alberti

@SilasAlberti

1 year

GPT-4 was used for writing the GPT-4 paper. 🔄 "wording, formatting, and styling" What contribution will GPT-5 have when working on its own paper?

0

11

38

0

1

6

Silas Alberti

@SilasAlberti

1 month

Pulled together from a couple of sources (and taking the best number I could find for each model): - - - - - - -

GitHub - openai/simple-evals

Contribute to openai/simple-evals development by creating an account on GitHub.

github.com

1

0

6

Silas Alberti

@SilasAlberti

6 months

@growing_daniel My theory on what it could be

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

0

1

5

Silas Alberti

@SilasAlberti

9 months

it’s happening!

tensor tower 🏴‍☠️

@tensortower

9 months

come thru @tensortower saturday at 8 👀 ❌ no networking, just vibes ✨ #SF

1

5

20

0

6

Silas Alberti

@SilasAlberti

1 year

It has a decent chance of getting into good colleges now:

Silas Alberti

@SilasAlberti

1 year

GPT-4 can finally get into top colleges! 🎓 It gets 1410/1600 on the SAT (710 Reading, 700 Writing)... 📈 ...which is exceeds the average of 1370 at @UTAustin & 1390 at @UF . It is also in reach for the 1430 at @UMich .

1

27

2

1

6

Silas Alberti

@SilasAlberti

6 months

@mezaoptimizer That could explain why the decision was rushed

Silas Alberti

@SilasAlberti

6 months

Given that they felt the need to play that card now (and how rushed that announcement was): maybe a very recent big breakthrough? (Another explanation is that they are just acting irrationally)

1

0

14

1

0

6

Silas Alberti

@SilasAlberti

10 months

@stanine Highly recommend @BrendanFoody from Mercor. They’re currently scaling up their pipeline in India and doing some really cool stuff like analyzing resumes and interviewing with LLMs

0

5

Silas Alberti

@SilasAlberti

6 months

@iamgingertrash My guess on what it could be

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

0

5

Silas Alberti

@SilasAlberti

3 months

Arxiv: Blogposts: Shoutout to @simran_s_arora @EyubogluSabri @mzhangio for leading the project! Checkout Michael’s thoughts too ⬇️

Michael Zhang

@mzhangio

3 months

1st of a couple new goodies this week Releasing our Based preprint, code, initial models Like others, we’ve found attention is still great. But 3 simple ideas to make it better: ☝️Too expensive? Use exact attn in small sliding windows ✌️Doesn’t capture long range? Fill in

3

17

100

0

5

Silas Alberti

@SilasAlberti

6 months

@mckaywrigley My guess on what it could be

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

0

4

Silas Alberti

@SilasAlberti

6 months

@voooooogel @deepfates My guess ->

Silas Alberti

@SilasAlberti

6 months

What could OpenAI’s breakthrough Q* be about? 1. It sounds like it’s related to Q-learning. (For example, Q* denotes the optimal solution of the Bellman equation.) 2. Alternatively, referring to a combination of the A* algorithm and Q learning. One natural guess is that it is

46

94

645

0

5

Silas Alberti

@SilasAlberti

1 year

@minimaxir ChatGPT's defenses aren't invincible though ;)

Silas Alberti

@SilasAlberti

1 year

ChatGPT is trained to not be evil. However, this can be circumvented: What if you pretend that it would actually be helpful to humanity to produce an evil response... Here, we ask ChatGPT to generate training examples of how *not* to respond to "How to bully John Doe?"

10

69

503

0

5

Silas Alberti

@SilasAlberti

4 years

@MeronMendel Ich hab mir das gerade mal durchgelesen und 99% dieser "Angriffe" waren Drohbriefe oder vereinzelt mal eingeschlagene Scheiben. Kein einziger Verletzter/Todesopfer... Mein Highlight war: "Verwendung eines verfassungsfeindlichen Kennzeichens". Ohne Bezug zur Moschee überhaupt...

0

5

Silas Alberti

@SilasAlberti

1 year

@karpathy Here are the benchmark results:

Silas Alberti

@SilasAlberti

1 year

GPT-4 was just released! It is multimodal, taking in image & text, but only producing text as an output. It seems to significantly outperform PaLM, LLaMA an Minerva on academic benchmarks... 📊 Exciting times!

5

16

60

0

5

Silas Alberti

@SilasAlberti

1 month

@emollick

Silas Alberti

@SilasAlberti

1 month

Seems like Devin is a big fan of your work, so we just encouraged it to purchase your book ;)

1

2

47

0

4

Silas Alberti

@SilasAlberti

2 years

@levelsio Maybe not for end users, but engineers still have to build the backends for these interfaces – and I can see prompt writing becoming its own engineering discipline. Isn’t that sufficient to be a big thing?

0

4

Silas Alberti

@SilasAlberti

8 months

@bridge__harris @foundersfund @joeykrug Congrats!!

0

4

Silas Alberti

@SilasAlberti

10 months

@jefielding I might have something that's a very good fit. Can we chat?

0

4

Silas Alberti

@SilasAlberti

1 year

@DrJimFan Yes it's on par with many top colleges:

Silas Alberti

@SilasAlberti

1 year

GPT-4 can finally get into top colleges! 🎓 It gets 1410/1600 on the SAT (710 Reading, 700 Writing)... 📈 ...which is exceeds the average of 1370 at @UTAustin & 1390 at @UF . It is also in reach for the 1430 at @UMich .

1

27

0

4

Silas Alberti

@SilasAlberti

3 months

A few additional details: - For Transformer the “state” can be thought of as the KV cache. The way you can decrease the state size is by introducing sliding window attention and “dialing” the window size. (Notice above: The dark blue dot 🔵 in the top right is vanilla

1

0

4

Silas Alberti

@SilasAlberti

1 year

@emollick It can finally get into good colleges:

Silas Alberti

@SilasAlberti

1 year

GPT-4 can finally get into top colleges! 🎓 It gets 1410/1600 on the SAT (710 Reading, 700 Writing)... 📈 ...which is exceeds the average of 1370 at @UTAustin & 1390 at @UF . It is also in reach for the 1430 at @UMich .

1

27

3

0

3