Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

I wrote my own CAD software whose intended user is an AI. I asked GPT 5.6 sol to make me a case for my camera. It only interrupted me once, which was to log in to a website to download a model of my camera as reference. Not perfect... show more

kache

389,993 subscribers

94,132 views • 1 month ago •via X (Twitter)

Science & Technology

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

me when someone asks to see my camera roll but i’m an artist so i have a heinous amount of reference pictures saved that make me look legitimately insane

me when someone asks to see my camera roll but i’m an artist so i have a heinous amount of reference pictures saved that make me look legitimately insane

kitty⋆˚꩜｡

27,505 views • 1 year ago

i just created an entire video with gpt-5.6 sol for site every scene, camera angle, animation - is a line of code in react zero traditional editing. gpt-5.6 and @remotion are legendary

i just created an entire video with gpt-5.6 sol for site every scene, camera angle, animation - is a line of code in react zero traditional editing. gpt-5.6 and @remotion are legendary

Suhail Kakar

27,207 views • 1 month ago

This is me fighting to save my camera at the Togolese border because I refused to pay a bribe of CFA 2000 in order for my passport to stamped😎

This is me fighting to save my camera at the Togolese border because I refused to pay a bribe of CFA 2000 in order for my passport to stamped😎

Wode Maya ®

1,614,505 views • 3 years ago

I know my camera lens is not clean, but it’s impossible for me right now to move and find a cloth, I somehow managed to get on wheelchair and right now I’m sitting next to the window..I thought of recording a video because i was feeling better.. i’ve been listening to the life of a show girl, so yeah, my mood is better and my energy feels restoring…cheers!

I know my camera lens is not clean, but it’s impossible for me right now to move and find a cloth, I somehow managed to get on wheelchair and right now I’m sitting next to the window..I thought of recording a video because i was feeling better.. i’ve been listening to the life of a show girl, so yeah, my mood is better and my energy feels restoring…cheers!

nasty plot

11,454 views • 6 months ago

I was under threat of a spanking. It's not my fault my diaper was so full that it was pushing my legs apart that's the only reason I fell down on my padded bottom! And I'm shaking my rattle because she handed it to me, and whose not going to shake a rattle? That's suspicious!

Sensitive content

I was under threat of a spanking. It's not my fault my diaper was so full that it was pushing my legs apart that's the only reason I fell down on my padded bottom! And I'm shaking my rattle because she handed it to me, and whose not going to shake a rattle? That's suspicious!

PrettyJessica

30,392 views • 1 year ago

i was jus walking around gbtb talking to my dad then this guy walked past me (mind u its dark as hell here) and he was talking to a camera n it didnt compute in my head right away it took a second for me to realise… i turn around n its literally seungjin with their managers like

i was jus walking around gbtb talking to my dad then this guy walked past me (mind u its dark as hell here) and he was talking to a camera n it didnt compute in my head right away it took a second for me to realise… i turn around n its literally seungjin with their managers like

SAW SKZ…

362,244 views • 3 years ago

AI is outrageously good. I built an AI digital twin of myself in 2 hours. I am not a developer. But I had an idea to add an AI chat to my website using only Cursor, Claude Code, and ElevenLabs (without Figma). Designers should build more. (take a look:

AI is outrageously good. I built an AI digital twin of myself in 2 hours. I am not a developer. But I had an idea to add an AI chat to my website using only Cursor, Claude Code, and ElevenLabs (without Figma). Designers should build more. (take a look:

Felix Lee

13,159 views • 9 months ago

back in my old house i was checking my camera to see if my cats were messing with my lizard who’s tank was in the kitchen and then i saw this light and heard those sounds it made and i threw my camera out. i tried so hard to debunk it bc i’m a scared of ghosts but couldn’t.

back in my old house i was checking my camera to see if my cats were messing with my lizard who’s tank was in the kitchen and then i saw this light and heard those sounds it made and i threw my camera out. i tried so hard to debunk it bc i’m a scared of ghosts but couldn’t.

jamie

33,105 views • 4 months ago

Not an AI account. Not a Russian bot. Not paid to push any content or an agenda. Not affiliated with any group. Not influenced by anyone. All of my thoughts, opinions & words are my own. I will not accept money to promote an agenda or push a narrative. This has caused me issues but I can’t be bought. I am just an Aussie girl who loves her country, her people, her culture & way of life. I will do whatever I need to, to protect it. As I have quite a few new followers, I thought it was only fair to clear this up & just yesterday an influencer referred to me as an AI account 🙄 I am a real human being & I’m not sure why this keeps coming up 🤷🏼‍♀️ 𝑳𝒐𝒛𝒛𝒚 𝑩 ♥️

Not an AI account. Not a Russian bot. Not paid to push any content or an agenda. Not affiliated with any group. Not influenced by anyone. All of my thoughts, opinions & words are my own. I will not accept money to promote an agenda or push a narrative. This has caused me issues but I can’t be bought. I am just an Aussie girl who loves her country, her people, her culture & way of life. I will do whatever I need to, to protect it. As I have quite a few new followers, I thought it was only fair to clear this up & just yesterday an influencer referred to me as an AI account 🙄 I am a real human being & I’m not sure why this keeps coming up 🤷🏼‍♀️ 𝑳𝒐𝒛𝒛𝒚 𝑩 ♥️

Lozzy B 🇦🇺𝕏

47,367 views • 11 months ago

During an interview, Stephen Kiki Ramos, A 17 years old Haiti Origin, Who just made his first debut as U17 player for Argentina, Was asked what it felt like Representing the Country, 🗣️ "I honestly can't even express how happy and proud I am right now. When I was first called up for training, a lot of my teammates didn't seem to like me because of my skin color. Some of them even mocked me to my face, calling me 'black monkey' and saying I'm not a real Argentine. It hurt I won't lie. I was sad and I called my mom. She encouraged me not to let it weigh me down. She told me more of that would come, but I should never let it get to me. And now? I think my teammates are finally starting to accept me. I'm so happy to make my debut, and I'm ready for the big journey ahead."

During an interview, Stephen Kiki Ramos, A 17 years old Haiti Origin, Who just made his first debut as U17 player for Argentina, Was asked what it felt like Representing the Country, 🗣️ "I honestly can't even express how happy and proud I am right now. When I was first called up for training, a lot of my teammates didn't seem to like me because of my skin color. Some of them even mocked me to my face, calling me 'black monkey' and saying I'm not a real Argentine. It hurt I won't lie. I was sad and I called my mom. She encouraged me not to let it weigh me down. She told me more of that would come, but I should never let it get to me. And now? I think my teammates are finally starting to accept me. I'm so happy to make my debut, and I'm ready for the big journey ahead."

Femzy💫

51,071 views • 1 month ago

So this is a security camera footage when my parents ran to congratulate me after I won we cried like crazy TwT

So this is a security camera footage when my parents ran to congratulate me after I won we cried like crazy TwT

GranTODAKAI

32,490 views • 1 year ago

First off no I’m not a dude and the ONLY person running my account is me. Secondly I will be online to drain you losers all night. So get in line and approach properly. I make my own posts, do my own messages, bully my own losers. I don’t need a beta bitch to do it because I actually enjoy it. So also stop messaging me that you want to manage my account bc you will “make me so much more money” when your some loser in a 3rd world country.

Sensitive content

First off no I’m not a dude and the ONLY person running my account is me. Secondly I will be online to drain you losers all night. So get in line and approach properly. I make my own posts, do my own messages, bully my own losers. I don’t need a beta bitch to do it because I actually enjoy it. So also stop messaging me that you want to manage my account bc you will “make me so much more money” when your some loser in a 3rd world country.

Princess Adri

23,835 views • 5 months ago

This is my "feel the AGI" moment: I used GPT-5.6 Sol to train my own autocorrect model that outperforms GPT-5.6 Sol (wtf??) I have no ML background. I have no idea what I'm doing. I just kept pushing Sol until it spat out a SOTA model. And I spent $0. The motivation: Years of talking to AI have made me terrible at typing. Rather than fix my skill issue, I decided to throw more AI at it. My idea was: instead of autocorrect that interrupts my flow, I want to type fast with mistakes and have AI clean it up after. I wanted the smallest local model possible, for speed, for battery life, for science! So I decided to train my own. Inspired by Andrej Karpathy’s autoresearch, I ran Codex /goal with this setup: pick an experiment, try it, record the results to a doc, throw it out if it fails, and plan the next experiment without repeating failures. I gave a few examples that had to pass, tight latency targets, and let it run. Sol did some amazing things. First, it scanned benchmarks and shortlisted base models: Qwen 3.5, Gemma 4, Liquid LFM 2.5. It found a dataset on HuggingFace for typed text. Then it built a simulator for fingers striking a Mac keyboard, modeling the physical layout with a Gaussian distribution around each key. It simulated striking the wrong key, wrong order, fat-fingering, etc. With the models + data + simulator, it fine-tuned using MLX right on my MacBook. It had a working prototype within an hour! But accuracy was pretty poor. — Problem 1: Tokenization Sol read papers, ran tests, and identified that the tokenizer was the bottleneck. Tokenization makes typos hard for the model to see, so it memorizes mappings instead of using its language priors. Sol tried ByT5, Google’s tokenizer-free byte-level LLM. This made a big improvement, but the model is old and lacked the knowledge needed to reach Sol performance. Sol dug deeper and realized a tokenizer-free model isn’t needed; instead, it used T5Gemma, an encoder-decoder model. This can understand the input deeply before producing output, and furthermore, Sol could post-train the encoder to improve performance. This gave a much higher ceiling. — Problem 2: Loss function Now the model was correcting some typos perfectly, but ignoring most. Sol realized that standard cross-entropy loss was teaching the model to avoid edits, because the vast majority of characters in the training data were left unmodified. The fix was wild: Sol wrote a custom loss function that byte-aligns the source and target strings, uses a dynamic programming algorithm to compute the minimum edits between the two, then weights correct edits much higher than copies. After a lot of tuning, this dramatically improved accuracy. — Problem 3: Autoregression One failure mode remained: if the model made a mistake, it couldn’t backtrack. It could only predict the next token. Teaching it to “think” like a reasoning model would solve this, but would be far too slow. Sol found a beautiful solution: instead of greedily predicting the next token, beam search over all possibilities. This parallelizes the exploration instead of one linear chain-of-thought. At the end, choose the path with highest cumulative log probability. This worked great, but made the experience worse, since the user wouldn’t see progress until the whole search was done. To fix this, Sol made a clever observation: after each search step, the longest common prefix among surviving branches is guaranteed to appear in the final result, so it can be displayed immediately. As the search progresses, weaker paths are dropped and the prefix grows, so the user sees continuous progress. Sol built all this as a custom MLX pipeline that does the parallel decoding on the MacBook GPU, with just ~40ms TTFT. It’s crazy fast and entirely local. — Final eval (error reduction rate, higher is better): - Apple autocorrect: 49.66% - GPT-5.6 Luna: 82.47% - GPT-5.6 Terra: 87.64% - GPT-5.6 Sol: 90.56% - Our model (1.7B): 91.02% Final cost: - 1 quota reset (thanks Tibo) - $0 (And yes, I verified there's no cheating. In fact, we test words scrubbed from the training data to prove the model isn’t memorizing) There were a ton more details and tangents I could write about: contrastive learning, GRPO, DPO, dynamic masking, and more. Sol is a fascinating and creative model. It blew my mind so many times. Don’t let a lack of experience stop you: Sol makes AI experiments accessible to anyone!

This is my "feel the AGI" moment: I used GPT-5.6 Sol to train my own autocorrect model that outperforms GPT-5.6 Sol (wtf??) I have no ML background. I have no idea what I'm doing. I just kept pushing Sol until it spat out a SOTA model. And I spent $0. The motivation: Years of talking to AI have made me terrible at typing. Rather than fix my skill issue, I decided to throw more AI at it. My idea was: instead of autocorrect that interrupts my flow, I want to type fast with mistakes and have AI clean it up after. I wanted the smallest local model possible, for speed, for battery life, for science! So I decided to train my own. Inspired by Andrej Karpathy’s autoresearch, I ran Codex /goal with this setup: pick an experiment, try it, record the results to a doc, throw it out if it fails, and plan the next experiment without repeating failures. I gave a few examples that had to pass, tight latency targets, and let it run. Sol did some amazing things. First, it scanned benchmarks and shortlisted base models: Qwen 3.5, Gemma 4, Liquid LFM 2.5. It found a dataset on HuggingFace for typed text. Then it built a simulator for fingers striking a Mac keyboard, modeling the physical layout with a Gaussian distribution around each key. It simulated striking the wrong key, wrong order, fat-fingering, etc. With the models + data + simulator, it fine-tuned using MLX right on my MacBook. It had a working prototype within an hour! But accuracy was pretty poor. — Problem 1: Tokenization Sol read papers, ran tests, and identified that the tokenizer was the bottleneck. Tokenization makes typos hard for the model to see, so it memorizes mappings instead of using its language priors. Sol tried ByT5, Google’s tokenizer-free byte-level LLM. This made a big improvement, but the model is old and lacked the knowledge needed to reach Sol performance. Sol dug deeper and realized a tokenizer-free model isn’t needed; instead, it used T5Gemma, an encoder-decoder model. This can understand the input deeply before producing output, and furthermore, Sol could post-train the encoder to improve performance. This gave a much higher ceiling. — Problem 2: Loss function Now the model was correcting some typos perfectly, but ignoring most. Sol realized that standard cross-entropy loss was teaching the model to avoid edits, because the vast majority of characters in the training data were left unmodified. The fix was wild: Sol wrote a custom loss function that byte-aligns the source and target strings, uses a dynamic programming algorithm to compute the minimum edits between the two, then weights correct edits much higher than copies. After a lot of tuning, this dramatically improved accuracy. — Problem 3: Autoregression One failure mode remained: if the model made a mistake, it couldn’t backtrack. It could only predict the next token. Teaching it to “think” like a reasoning model would solve this, but would be far too slow. Sol found a beautiful solution: instead of greedily predicting the next token, beam search over all possibilities. This parallelizes the exploration instead of one linear chain-of-thought. At the end, choose the path with highest cumulative log probability. This worked great, but made the experience worse, since the user wouldn’t see progress until the whole search was done. To fix this, Sol made a clever observation: after each search step, the longest common prefix among surviving branches is guaranteed to appear in the final result, so it can be displayed immediately. As the search progresses, weaker paths are dropped and the prefix grows, so the user sees continuous progress. Sol built all this as a custom MLX pipeline that does the parallel decoding on the MacBook GPU, with just ~40ms TTFT. It’s crazy fast and entirely local. — Final eval (error reduction rate, higher is better): - Apple autocorrect: 49.66% - GPT-5.6 Luna: 82.47% - GPT-5.6 Terra: 87.64% - GPT-5.6 Sol: 90.56% - Our model (1.7B): 91.02% Final cost: - 1 quota reset (thanks Tibo) - $0 (And yes, I verified there's no cheating. In fact, we test words scrubbed from the training data to prove the model isn’t memorizing) There were a ton more details and tangents I could write about: contrastive learning, GRPO, DPO, dynamic masking, and more. Sol is a fascinating and creative model. It blew my mind so many times. Don’t let a lack of experience stop you: Sol makes AI experiments accessible to anyone!

Anshu

179,389 views • 1 month ago

Went from 0.3 SOL to 17 SOL in less than an hour. People told me to give up. Told me to log off. Told me I'd never make it. I'm still here. I'm going to keep fighting for my mom. F*** all the haters. And shoutout to sling for saving my port. I Know you don't like me but im grateful af because this money aint even for me.

Went from 0.3 SOL to 17 SOL in less than an hour. People told me to give up. Told me to log off. Told me I'd never make it. I'm still here. I'm going to keep fighting for my mom. F*** all the haters. And shoutout to sling for saving my port. I Know you don't like me but im grateful af because this money aint even for me.

Nofe

207,288 views • 23 days ago

No one believed me when I said that my cat is a bully, so l setup a camera to prove it...

No one believed me when I said that my cat is a bully, so l setup a camera to prove it...

We don't deserve cats 😺

1,346,556 views • 9 months ago

Codex + 5.5 + computer use built me an iPhone shortcut to make it easy to use GPT Image 2 on any photo in my camera roll. Then it edited this screen recording with ffmpeg to speed it up so I could share it. Great for making quick UI mocks on my phone. And for lulz.

Codex + 5.5 + computer use built me an iPhone shortcut to make it easy to use GPT Image 2 on any photo in my camera roll. Then it edited this screen recording with ffmpeg to speed it up so I could share it. Great for making quick UI mocks on my phone. And for lulz.

Ilya Abyzov

45,451 views • 4 months ago

So gym trade just told me how he had a hidden camera which explains why he rushed me to get on my underwear/naked… idk I just feel so stupid like I knew he was just interested in me because of my body and how innocent I was for a 23 year old…

Sensitive content

So gym trade just told me how he had a hidden camera which explains why he rushed me to get on my underwear/naked… idk I just feel so stupid like I knew he was just interested in me because of my body and how innocent I was for a 23 year old…

Anne 🎀

17,439 views • 3 months ago

🚨 Qwen new checkpoint - kieran As usual it claims he is claude and made by anthropic but with tokenizer prompt , its clear - QWEN model Ignore website this is my test to know which model is which behind the scene , should i test it and compare with kimi k3 / opus 5 /sol 5.6

🚨 Qwen new checkpoint - kieran As usual it claims he is claude and made by anthropic but with tokenizer prompt , its clear - QWEN model Ignore website this is my test to know which model is which behind the scene , should i test it and compare with kimi k3 / opus 5 /sol 5.6

Chetaslua

15,544 views • 1 month ago

Throwback to 2021: my friends got me a Raja cameo for my birthday. I was in a local drag comp and they asked her to encourage me andshe thought it was drag race 🤣. Skip to 25 and I got to get booted by my idol 😌😌

Throwback to 2021: my friends got me a Raja cameo for my birthday. I was in a local drag comp and they asked her to encourage me andshe thought it was drag race 🤣. Skip to 25 and I got to get booted by my idol 😌😌

Ivory Glaze

107,496 views • 7 months ago