Codedidit
@ogtzor.bsky.social
What’s got you most excited today?
March 4, 2025 at 9:39 PM
What’s got you most excited today?
Your Customer Persona Is a Lie and It’s Costing You Millions
Your Customers Don’t Actually Care About What You Think They Do → You’re selling features, but they’re buying outcomes. Do you know what they actually want?
Your Customers Don’t Actually Care About What You Think They Do → You’re selling features, but they’re buying outcomes. Do you know what they actually want?
March 3, 2025 at 12:54 AM
Your Customer Persona Is a Lie and It’s Costing You Millions
Your Customers Don’t Actually Care About What You Think They Do → You’re selling features, but they’re buying outcomes. Do you know what they actually want?
Your Customers Don’t Actually Care About What You Think They Do → You’re selling features, but they’re buying outcomes. Do you know what they actually want?
Your Marketing Isn’t the Problem, Your Product Is Just Forgettable.
Your Users Don’t Know How to Explain Your Product to Others → If your product isn’t simple to describe, it won’t spread.
Viral products aren’t just great, they’re easy to talk about.
Your Users Don’t Know How to Explain Your Product to Others → If your product isn’t simple to describe, it won’t spread.
Viral products aren’t just great, they’re easy to talk about.
March 1, 2025 at 9:53 PM
Your Marketing Isn’t the Problem, Your Product Is Just Forgettable.
Your Users Don’t Know How to Explain Your Product to Others → If your product isn’t simple to describe, it won’t spread.
Viral products aren’t just great, they’re easy to talk about.
Your Users Don’t Know How to Explain Your Product to Others → If your product isn’t simple to describe, it won’t spread.
Viral products aren’t just great, they’re easy to talk about.
CodeGemma from Google demonstrates how specialized AI models can transform development workflows.
The implementation uses Flash Attention 2 for optimized performance, making it practical for production environments.
The implementation uses Flash Attention 2 for optimized performance, making it practical for production environments.
March 1, 2025 at 10:04 AM
CodeGemma from Google demonstrates how specialized AI models can transform development workflows.
The implementation uses Flash Attention 2 for optimized performance, making it practical for production environments.
The implementation uses Flash Attention 2 for optimized performance, making it practical for production environments.
MLGym is here to push AI research forward.
This new framework benchmarks LLM agents on reinforcement learning tasks, helping researchers compare models and accelerate progress.
This new framework benchmarks LLM agents on reinforcement learning tasks, helping researchers compare models and accelerate progress.
February 28, 2025 at 10:57 PM
MLGym is here to push AI research forward.
This new framework benchmarks LLM agents on reinforcement learning tasks, helping researchers compare models and accelerate progress.
This new framework benchmarks LLM agents on reinforcement learning tasks, helping researchers compare models and accelerate progress.
SigLIP 2 takes vision-language AI to the next level.
With better semantic understanding and localization, it excels across languages.
With better semantic understanding and localization, it excels across languages.
February 28, 2025 at 6:01 PM
SigLIP 2 takes vision-language AI to the next level.
With better semantic understanding and localization, it excels across languages.
With better semantic understanding and localization, it excels across languages.
Better code from AI? This hybrid test-time scaling framework boosts accuracy and coverage in code generation.
It might use dynamic scaling during inference to adapt outputs to specific inputs, reducing bugs and edge-case failures. Developers, this could change your workflow.
It might use dynamic scaling during inference to adapt outputs to specific inputs, reducing bugs and edge-case failures. Developers, this could change your workflow.
February 27, 2025 at 8:01 PM
Better code from AI? This hybrid test-time scaling framework boosts accuracy and coverage in code generation.
It might use dynamic scaling during inference to adapt outputs to specific inputs, reducing bugs and edge-case failures. Developers, this could change your workflow.
It might use dynamic scaling during inference to adapt outputs to specific inputs, reducing bugs and edge-case failures. Developers, this could change your workflow.
Updating LLMs without breaking them?
This paper explores LoRA adapters for stuffing new facts into models efficiently.
LoRA tweaks only a few parameters to avoid catastrophic forgetting, but the trick is balancing old and new knowledge. Key for keeping AI current.
This paper explores LoRA adapters for stuffing new facts into models efficiently.
LoRA tweaks only a few parameters to avoid catastrophic forgetting, but the trick is balancing old and new knowledge. Key for keeping AI current.
February 27, 2025 at 3:03 PM
Updating LLMs without breaking them?
This paper explores LoRA adapters for stuffing new facts into models efficiently.
LoRA tweaks only a few parameters to avoid catastrophic forgetting, but the trick is balancing old and new knowledge. Key for keeping AI current.
This paper explores LoRA adapters for stuffing new facts into models efficiently.
LoRA tweaks only a few parameters to avoid catastrophic forgetting, but the trick is balancing old and new knowledge. Key for keeping AI current.
SuperGPQA tests LLMs across 285 graduate fields, revealing their strengths and gaps. From math to medicine, it’s a reality check for AI knowledge.
It probably uses complex reasoning tasks to probe deep comprehension, not just rote answers.
Curious about LLM limits? This has the answers.
It probably uses complex reasoning tasks to probe deep comprehension, not just rote answers.
Curious about LLM limits? This has the answers.
February 27, 2025 at 1:00 AM
SuperGPQA tests LLMs across 285 graduate fields, revealing their strengths and gaps. From math to medicine, it’s a reality check for AI knowledge.
It probably uses complex reasoning tasks to probe deep comprehension, not just rote answers.
Curious about LLM limits? This has the answers.
It probably uses complex reasoning tasks to probe deep comprehension, not just rote answers.
Curious about LLM limits? This has the answers.
AI is revolutionizing healthcare with machine learning algorithms that crunch medical data, like spotting cancer in X-rays.
It also uses ensemble methods, combining multiple models for sharper predictions, helping docs catch diseases early and tweak treatments.
It also uses ensemble methods, combining multiple models for sharper predictions, helping docs catch diseases early and tweak treatments.
February 26, 2025 at 5:59 PM
AI is revolutionizing healthcare with machine learning algorithms that crunch medical data, like spotting cancer in X-rays.
It also uses ensemble methods, combining multiple models for sharper predictions, helping docs catch diseases early and tweak treatments.
It also uses ensemble methods, combining multiple models for sharper predictions, helping docs catch diseases early and tweak treatments.
Google's CodeGemma represents a significant advancement in AI-assisted development.
Available in three variants (Base, Instruct, and It), it's specifically optimized for different coding scenarios from general tasks to IDE integration.
Available in three variants (Base, Instruct, and It), it's specifically optimized for different coding scenarios from general tasks to IDE integration.
February 26, 2025 at 2:04 PM
Google's CodeGemma represents a significant advancement in AI-assisted development.
Available in three variants (Base, Instruct, and It), it's specifically optimized for different coding scenarios from general tasks to IDE integration.
Available in three variants (Base, Instruct, and It), it's specifically optimized for different coding scenarios from general tasks to IDE integration.
OpenAI’s o3-mini is a game-changer for STEM.
It’s built on a transformer model, the backbone of modern language AIs, stacking layers of “attention” to juggle complex tasks like coding or science questions.
It’s built on a transformer model, the backbone of modern language AIs, stacking layers of “attention” to juggle complex tasks like coding or science questions.
February 25, 2025 at 7:59 PM
OpenAI’s o3-mini is a game-changer for STEM.
It’s built on a transformer model, the backbone of modern language AIs, stacking layers of “attention” to juggle complex tasks like coding or science questions.
It’s built on a transformer model, the backbone of modern language AIs, stacking layers of “attention” to juggle complex tasks like coding or science questions.
Just explored Google's CodeGemma .
An impressive open-weight model for code generation in 7B and 2B parameter sizes. Its 8,192 token context window enables handling complex coding tasks with remarkable efficiency.
An impressive open-weight model for code generation in 7B and 2B parameter sizes. Its 8,192 token context window enables handling complex coding tasks with remarkable efficiency.
February 25, 2025 at 4:56 PM
Just explored Google's CodeGemma .
An impressive open-weight model for code generation in 7B and 2B parameter sizes. Its 8,192 token context window enables handling complex coding tasks with remarkable efficiency.
An impressive open-weight model for code generation in 7B and 2B parameter sizes. Its 8,192 token context window enables handling complex coding tasks with remarkable efficiency.
2025 is AI’s big year.
We’re seeing more efficient models, think model distillation, shrinking big AIs into speedy little ones, and wild mashups like AI with quantum computing to solve problems crazy fast.
From gaming to healthcare, it’s endless.
We’re seeing more efficient models, think model distillation, shrinking big AIs into speedy little ones, and wild mashups like AI with quantum computing to solve problems crazy fast.
From gaming to healthcare, it’s endless.
February 25, 2025 at 3:01 PM
2025 is AI’s big year.
We’re seeing more efficient models, think model distillation, shrinking big AIs into speedy little ones, and wild mashups like AI with quantum computing to solve problems crazy fast.
From gaming to healthcare, it’s endless.
We’re seeing more efficient models, think model distillation, shrinking big AIs into speedy little ones, and wild mashups like AI with quantum computing to solve problems crazy fast.
From gaming to healthcare, it’s endless.
Autonomous AI agents are here to save the day.
They use reinforcement learning, trial-and-error with a reward twist, or planning algorithms like searching a decision tree to pick your next trip or book.
Some even tap multi-agent systems, teaming up with other AIs to tackle bigger tasks.
They use reinforcement learning, trial-and-error with a reward twist, or planning algorithms like searching a decision tree to pick your next trip or book.
Some even tap multi-agent systems, teaming up with other AIs to tackle bigger tasks.
February 25, 2025 at 7:58 AM
Autonomous AI agents are here to save the day.
They use reinforcement learning, trial-and-error with a reward twist, or planning algorithms like searching a decision tree to pick your next trip or book.
Some even tap multi-agent systems, teaming up with other AIs to tackle bigger tasks.
They use reinforcement learning, trial-and-error with a reward twist, or planning algorithms like searching a decision tree to pick your next trip or book.
Some even tap multi-agent systems, teaming up with other AIs to tackle bigger tasks.
Minecraft just got wilder with AI. Tools like Oasis use convolutional neural networks (CNNs), think of them as AI’s pattern-spotting superpowers, to generate infinite, unpredictable worlds.
February 24, 2025 at 9:04 PM
Minecraft just got wilder with AI. Tools like Oasis use convolutional neural networks (CNNs), think of them as AI’s pattern-spotting superpowers, to generate infinite, unpredictable worlds.
For me Grok 3 has been smart, creative, and fun.
Its ‘Think’ button uses interpretability techniques, like tracing the AI’s decision paths, to show how it reasons, almost like a behind-the-scenes tour.
Its ‘Think’ button uses interpretability techniques, like tracing the AI’s decision paths, to show how it reasons, almost like a behind-the-scenes tour.
February 24, 2025 at 5:00 PM
For me Grok 3 has been smart, creative, and fun.
Its ‘Think’ button uses interpretability techniques, like tracing the AI’s decision paths, to show how it reasons, almost like a behind-the-scenes tour.
Its ‘Think’ button uses interpretability techniques, like tracing the AI’s decision paths, to show how it reasons, almost like a behind-the-scenes tour.
When self-doubt and excuses creep in, remember this:
every time you tell yourself “I can’t,” you’re essentially saying “I don’t trust myself enough to learn what I don’t know yet.”
every time you tell yourself “I can’t,” you’re essentially saying “I don’t trust myself enough to learn what I don’t know yet.”
February 24, 2025 at 1:28 PM
When self-doubt and excuses creep in, remember this:
every time you tell yourself “I can’t,” you’re essentially saying “I don’t trust myself enough to learn what I don’t know yet.”
every time you tell yourself “I can’t,” you’re essentially saying “I don’t trust myself enough to learn what I don’t know yet.”
Education is broken. Not because teachers aren’t trying, not because students aren’t capable.
But because the system haven't evolved.
1.2 billion students worldwide are affected by learning disruptions.
40% of high school students say they feel disengaged in class.
But because the system haven't evolved.
1.2 billion students worldwide are affected by learning disruptions.
40% of high school students say they feel disengaged in class.
February 24, 2025 at 12:02 PM
Education is broken. Not because teachers aren’t trying, not because students aren’t capable.
But because the system haven't evolved.
1.2 billion students worldwide are affected by learning disruptions.
40% of high school students say they feel disengaged in class.
But because the system haven't evolved.
1.2 billion students worldwide are affected by learning disruptions.
40% of high school students say they feel disengaged in class.
DOOM fans, this is for you.
GameNGen uses generative adversarial networks (GANs), where one AI creates levels and another critiques them until they’re spot-on, to make DOOM worlds so real, you’d swear they’re human-made.
GameNGen uses generative adversarial networks (GANs), where one AI creates levels and another critiques them until they’re spot-on, to make DOOM worlds so real, you’d swear they’re human-made.
February 23, 2025 at 8:03 PM
DOOM fans, this is for you.
GameNGen uses generative adversarial networks (GANs), where one AI creates levels and another critiques them until they’re spot-on, to make DOOM worlds so real, you’d swear they’re human-made.
GameNGen uses generative adversarial networks (GANs), where one AI creates levels and another critiques them until they’re spot-on, to make DOOM worlds so real, you’d swear they’re human-made.
Imagine exploring a 3D game world crafted entirely by AI.
Google DeepMind's Genie 2 uses reinforcement learning, where the AI learns by trying things out and getting “rewards” for good moves, to build interactive environments from just one image prompt.
Google DeepMind's Genie 2 uses reinforcement learning, where the AI learns by trying things out and getting “rewards” for good moves, to build interactive environments from just one image prompt.
February 23, 2025 at 1:56 PM
Imagine exploring a 3D game world crafted entirely by AI.
Google DeepMind's Genie 2 uses reinforcement learning, where the AI learns by trying things out and getting “rewards” for good moves, to build interactive environments from just one image prompt.
Google DeepMind's Genie 2 uses reinforcement learning, where the AI learns by trying things out and getting “rewards” for good moves, to build interactive environments from just one image prompt.
Big data, small budget? Analyzing 10B tweets seemed impossible with commercial API costs. 💸 This study shows fine-tuned open-source models run locally & match GPT-4, making large-scale research actually feasible!
February 23, 2025 at 8:54 AM
Big data, small budget? Analyzing 10B tweets seemed impossible with commercial API costs. 💸 This study shows fine-tuned open-source models run locally & match GPT-4, making large-scale research actually feasible!
Thread time.
Imagine an AI that reads thousands of research papers overnight and hands you the highlights by breakfast. Google Research just dropped a game-changer: an AI co-scientist built to speed up breakthroughs. Here’s how it works.
Source: research.google/blog/accele...
Imagine an AI that reads thousands of research papers overnight and hands you the highlights by breakfast. Google Research just dropped a game-changer: an AI co-scientist built to speed up breakthroughs. Here’s how it works.
Source: research.google/blog/accele...
Accelerating scientific breakthroughs with an AI co-scientist
research.google
February 21, 2025 at 6:03 PM
Thread time.
Imagine an AI that reads thousands of research papers overnight and hands you the highlights by breakfast. Google Research just dropped a game-changer: an AI co-scientist built to speed up breakthroughs. Here’s how it works.
Source: research.google/blog/accele...
Imagine an AI that reads thousands of research papers overnight and hands you the highlights by breakfast. Google Research just dropped a game-changer: an AI co-scientist built to speed up breakthroughs. Here’s how it works.
Source: research.google/blog/accele...
AI co-scientist’s hypotheses are evolving!
Check out the Elo growth (1300–1700) vs. baselines like DeepSeek & humans over time. Top-10 avg shows steady improvement
Check out the Elo growth (1300–1700) vs. baselines like DeepSeek & humans over time. Top-10 avg shows steady improvement
February 21, 2025 at 2:59 PM
AI co-scientist’s hypotheses are evolving!
Check out the Elo growth (1300–1700) vs. baselines like DeepSeek & humans over time. Top-10 avg shows steady improvement
Check out the Elo growth (1300–1700) vs. baselines like DeepSeek & humans over time. Top-10 avg shows steady improvement
Wow, this AI co-scientist system is wild! Scientists team up with agents that debate, rank hypotheses, & evolve ideas.
Using tools like web searches & memory.
Could this reshape research?
Check the design:
Using tools like web searches & memory.
Could this reshape research?
Check the design:
February 20, 2025 at 9:40 PM
Wow, this AI co-scientist system is wild! Scientists team up with agents that debate, rank hypotheses, & evolve ideas.
Using tools like web searches & memory.
Could this reshape research?
Check the design:
Using tools like web searches & memory.
Could this reshape research?
Check the design: