Yonatan Lavy
@yonatanlavy.bsky.social
I build stuff for you to build stuff easier
Thanks Sonnet 4.5
October 4, 2025 at 7:49 AM
Thanks Sonnet 4.5
Okay, GPT-5 is live. But is it the game-changer we all expected?
After an initial dive into the specs, the model, and the API, here's my take: The real story isn't one single feature, but the entire package.
🧵 of my thoughts below 👇
After an initial dive into the specs, the model, and the API, here's my take: The real story isn't one single feature, but the entire package.
🧵 of my thoughts below 👇
August 7, 2025 at 6:28 PM
Okay, GPT-5 is live. But is it the game-changer we all expected?
After an initial dive into the specs, the model, and the API, here's my take: The real story isn't one single feature, but the entire package.
🧵 of my thoughts below 👇
After an initial dive into the specs, the model, and the API, here's my take: The real story isn't one single feature, but the entire package.
🧵 of my thoughts below 👇
GPT 5 is the same level as Opus 4.1 (on SWE bench)
@OpenAI and @AnthropicAI seems to stay near the top together, this is amazing news that top competitors are this close to each other
@OpenAI and @AnthropicAI seems to stay near the top together, this is amazing news that top competitors are this close to each other
August 7, 2025 at 5:14 PM
GPT 5 is the same level as Opus 4.1 (on SWE bench)
@OpenAI and @AnthropicAI seems to stay near the top together, this is amazing news that top competitors are this close to each other
@OpenAI and @AnthropicAI seems to stay near the top together, this is amazing news that top competitors are this close to each other
Claude Code is dancing for me
What does your AI IDE do for you?
What does your AI IDE do for you?
June 24, 2025 at 3:00 PM
Claude Code is dancing for me
What does your AI IDE do for you?
What does your AI IDE do for you?
Fine, I'll build it myself
May 28, 2025 at 6:41 PM
Fine, I'll build it myself
I love that I'm able to have 3 agents running in parallel on my codebase in cursor
Each has it's own agent mode, planning, executing or critiquing the execution of major features.
For planning I still use dotallio, but the code itself, cursor is hands down the best
Each has it's own agent mode, planning, executing or critiquing the execution of major features.
For planning I still use dotallio, but the code itself, cursor is hands down the best
May 20, 2025 at 2:00 PM
I love that I'm able to have 3 agents running in parallel on my codebase in cursor
Each has it's own agent mode, planning, executing or critiquing the execution of major features.
For planning I still use dotallio, but the code itself, cursor is hands down the best
Each has it's own agent mode, planning, executing or critiquing the execution of major features.
For planning I still use dotallio, but the code itself, cursor is hands down the best
3/
Now let's look at the knowledge cutoff date.
If you access the "latest" 4o model via the api - it shows you "April 2024"
But accessing the 4o model via chat - shows "June 2024" - this is the same as 4.1!
Now let's look at the knowledge cutoff date.
If you access the "latest" 4o model via the api - it shows you "April 2024"
But accessing the 4o model via chat - shows "June 2024" - this is the same as 4.1!
April 15, 2025 at 2:04 PM
3/
Now let's look at the knowledge cutoff date.
If you access the "latest" 4o model via the api - it shows you "April 2024"
But accessing the 4o model via chat - shows "June 2024" - this is the same as 4.1!
Now let's look at the knowledge cutoff date.
If you access the "latest" 4o model via the api - it shows you "April 2024"
But accessing the 4o model via chat - shows "June 2024" - this is the same as 4.1!
2/
In the benchmark above it shows 4.1 scoring about 66% on GPQA diamond.
I've found one match for GPQA benchmark for the updated 4o 2025-march -
AND IT MATCHES EXACTLY. So the 4.1 model shows the same performance as the march 2025 gpt-4o.
(can see more in the link at the end)
In the benchmark above it shows 4.1 scoring about 66% on GPQA diamond.
I've found one match for GPQA benchmark for the updated 4o 2025-march -
AND IT MATCHES EXACTLY. So the 4.1 model shows the same performance as the march 2025 gpt-4o.
(can see more in the link at the end)
April 15, 2025 at 2:04 PM
2/
In the benchmark above it shows 4.1 scoring about 66% on GPQA diamond.
I've found one match for GPQA benchmark for the updated 4o 2025-march -
AND IT MATCHES EXACTLY. So the 4.1 model shows the same performance as the march 2025 gpt-4o.
(can see more in the link at the end)
In the benchmark above it shows 4.1 scoring about 66% on GPQA diamond.
I've found one match for GPQA benchmark for the updated 4o 2025-march -
AND IT MATCHES EXACTLY. So the 4.1 model shows the same performance as the march 2025 gpt-4o.
(can see more in the link at the end)
1/
OpenAI claims massive improvements on GPT-4.1 over the GPT-4o - but in the small details you can see in the sub header it says "2024-11-20".
But they've updated their 4o just recently, along with when the image generation came out.
Let's dig in deeper..
OpenAI claims massive improvements on GPT-4.1 over the GPT-4o - but in the small details you can see in the sub header it says "2024-11-20".
But they've updated their 4o just recently, along with when the image generation came out.
Let's dig in deeper..
April 15, 2025 at 2:04 PM
1/
OpenAI claims massive improvements on GPT-4.1 over the GPT-4o - but in the small details you can see in the sub header it says "2024-11-20".
But they've updated their 4o just recently, along with when the image generation came out.
Let's dig in deeper..
OpenAI claims massive improvements on GPT-4.1 over the GPT-4o - but in the small details you can see in the sub header it says "2024-11-20".
But they've updated their 4o just recently, along with when the image generation came out.
Let's dig in deeper..
OpenAI just lied.
They "Launched" GPT 4.1, even though it appears to be the SAME EXACT MODEL as the recently launched updated 4o on March.
Let me show you the exact details 🧵👇
They "Launched" GPT 4.1, even though it appears to be the SAME EXACT MODEL as the recently launched updated 4o on March.
Let me show you the exact details 🧵👇
April 15, 2025 at 2:04 PM
OpenAI just lied.
They "Launched" GPT 4.1, even though it appears to be the SAME EXACT MODEL as the recently launched updated 4o on March.
Let me show you the exact details 🧵👇
They "Launched" GPT 4.1, even though it appears to be the SAME EXACT MODEL as the recently launched updated 4o on March.
Let me show you the exact details 🧵👇
Watching three Cursor agents write my app in parallel…
Felt like witnessing the start of AGI
Cursor didn’t just help me code faster,
it made me an AI agent and multiplied my work 3x.
Felt like witnessing the start of AGI
Cursor didn’t just help me code faster,
it made me an AI agent and multiplied my work 3x.
April 14, 2025 at 11:07 AM
Watching three Cursor agents write my app in parallel…
Felt like witnessing the start of AGI
Cursor didn’t just help me code faster,
it made me an AI agent and multiplied my work 3x.
Felt like witnessing the start of AGI
Cursor didn’t just help me code faster,
it made me an AI agent and multiplied my work 3x.
4/
This isn’t a moonshot.
OpenAI’s own roadmap lays it out:
Chatbots → Reasoners → Agents (We are here) → Innovators → Organizations
👇
This isn’t a moonshot.
OpenAI’s own roadmap lays it out:
Chatbots → Reasoners → Agents (We are here) → Innovators → Organizations
👇
April 13, 2025 at 9:53 AM
4/
This isn’t a moonshot.
OpenAI’s own roadmap lays it out:
Chatbots → Reasoners → Agents (We are here) → Innovators → Organizations
👇
This isn’t a moonshot.
OpenAI’s own roadmap lays it out:
Chatbots → Reasoners → Agents (We are here) → Innovators → Organizations
👇
The next dev joining your team… might be an AI agent.
Not a copilot. Not a helper.
A full SWE agent that submits PRs, runs tests, and ships features.
OpenAI just quietly confirmed it. 👇
🧵
Not a copilot. Not a helper.
A full SWE agent that submits PRs, runs tests, and ships features.
OpenAI just quietly confirmed it. 👇
🧵
April 13, 2025 at 9:53 AM
The next dev joining your team… might be an AI agent.
Not a copilot. Not a helper.
A full SWE agent that submits PRs, runs tests, and ships features.
OpenAI just quietly confirmed it. 👇
🧵
Not a copilot. Not a helper.
A full SWE agent that submits PRs, runs tests, and ships features.
OpenAI just quietly confirmed it. 👇
🧵
April 7, 2025 at 7:07 PM
Is this for real? gpt 5 open source ?!?
April 1, 2025 at 8:18 AM
Is this for real? gpt 5 open source ?!?
I asked chatgpt to add a way to chat with your outie whilst you are refining at MDR
March 31, 2025 at 9:57 PM
I asked chatgpt to add a way to chat with your outie whilst you are refining at MDR
You've heard of studio ghibli-fying photos, but have you heard of selfy-fying paintings?
March 30, 2025 at 7:51 PM
You've heard of studio ghibli-fying photos, but have you heard of selfy-fying paintings?
what they do with all those new gpt-4o image subscription money
March 29, 2025 at 9:46 PM
what they do with all those new gpt-4o image subscription money
It do be like that
March 29, 2025 at 9:00 PM
It do be like that
look, I made a guide to gpt-4o image gen
March 26, 2025 at 9:58 PM
look, I made a guide to gpt-4o image gen
New keyboard, who dis?
March 26, 2025 at 8:09 AM
New keyboard, who dis?
How I feel like when I vibe code a feature in 5 minutes
March 21, 2025 at 9:59 AM
How I feel like when I vibe code a feature in 5 minutes
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
March 19, 2025 at 11:41 AM
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach
Let's try a different approach