ErwinH
@erwinlearns.bsky.social
Building the future of LLMs @ Google
Fine-tuning and LLMOps
My views
https://www.linkedin.com/in/erwinhuizenga/
Fine-tuning and LLMOps
My views
https://www.linkedin.com/in/erwinhuizenga/
DeepSeek has released their upgraded DeepSeek-R1-0528 model.
🔹 Improved benchmark performance
🔹Enhanced front-end capabilities
🔹Supports JSON output & function calling
We've made sure its available from day one on Vertex AI for deployment. with Nvidia's H200
🔹 Improved benchmark performance
🔹Enhanced front-end capabilities
🔹Supports JSON output & function calling
We've made sure its available from day one on Vertex AI for deployment. with Nvidia's H200
May 30, 2025 at 9:51 AM
DeepSeek has released their upgraded DeepSeek-R1-0528 model.
🔹 Improved benchmark performance
🔹Enhanced front-end capabilities
🔹Supports JSON output & function calling
We've made sure its available from day one on Vertex AI for deployment. with Nvidia's H200
🔹 Improved benchmark performance
🔹Enhanced front-end capabilities
🔹Supports JSON output & function calling
We've made sure its available from day one on Vertex AI for deployment. with Nvidia's H200
Are you running an AI startup? Google launched the AI Futures Fund.
Being part of this program will give startups early access to Google's latest AI models as well as resources, technical expertise and equity funding to accelerate their progress.
Apply here:
labs.google/aifuturesfund
Being part of this program will give startups early access to Google's latest AI models as well as resources, technical expertise and equity funding to accelerate their progress.
Apply here:
labs.google/aifuturesfund
May 26, 2025 at 3:19 AM
Are you running an AI startup? Google launched the AI Futures Fund.
Being part of this program will give startups early access to Google's latest AI models as well as resources, technical expertise and equity funding to accelerate their progress.
Apply here:
labs.google/aifuturesfund
Being part of this program will give startups early access to Google's latest AI models as well as resources, technical expertise and equity funding to accelerate their progress.
Apply here:
labs.google/aifuturesfund
Big A2A and ADK news from Google I/O. 📢
Today, we're announcing:
1️⃣An A2A Python SDK AND an updated A2A protocol spec (v0.2).
2️⃣A new Java SDK for Agent Development Kit (ADK) AND the v1.0.0 stable release of our Python ADK.
Read more in our launch blog 👇
developers.googleblog.com/en/agents-ad...
Today, we're announcing:
1️⃣An A2A Python SDK AND an updated A2A protocol spec (v0.2).
2️⃣A new Java SDK for Agent Development Kit (ADK) AND the v1.0.0 stable release of our Python ADK.
Read more in our launch blog 👇
developers.googleblog.com/en/agents-ad...
May 20, 2025 at 11:27 PM
Big A2A and ADK news from Google I/O. 📢
Today, we're announcing:
1️⃣An A2A Python SDK AND an updated A2A protocol spec (v0.2).
2️⃣A new Java SDK for Agent Development Kit (ADK) AND the v1.0.0 stable release of our Python ADK.
Read more in our launch blog 👇
developers.googleblog.com/en/agents-ad...
Today, we're announcing:
1️⃣An A2A Python SDK AND an updated A2A protocol spec (v0.2).
2️⃣A new Java SDK for Agent Development Kit (ADK) AND the v1.0.0 stable release of our Python ADK.
Read more in our launch blog 👇
developers.googleblog.com/en/agents-ad...
Don't miss the GoogleIO 2025 Keynotes! 🚀
I'm excited about our announcements. Watch the livestream to stay ahead of the curve.
Link below 👇
I'm excited about our announcements. Watch the livestream to stay ahead of the curve.
Link below 👇
May 20, 2025 at 3:27 AM
Don't miss the GoogleIO 2025 Keynotes! 🚀
I'm excited about our announcements. Watch the livestream to stay ahead of the curve.
Link below 👇
I'm excited about our announcements. Watch the livestream to stay ahead of the curve.
Link below 👇
Yesterday the DeepMind team introduced AlphaEvolve: a Gemini-powered coding agent for algorithm discovery.
What I find impressive is that a single system can contribute to numerous open problems in maths and computer science. Which proves that Agents can discover novel and useful algorithms
What I find impressive is that a single system can contribute to numerous open problems in maths and computer science. Which proves that Agents can discover novel and useful algorithms
May 16, 2025 at 1:40 AM
Yesterday the DeepMind team introduced AlphaEvolve: a Gemini-powered coding agent for algorithm discovery.
What I find impressive is that a single system can contribute to numerous open problems in maths and computer science. Which proves that Agents can discover novel and useful algorithms
What I find impressive is that a single system can contribute to numerous open problems in maths and computer science. Which proves that Agents can discover novel and useful algorithms
Google is running an Agents hackathon using the Agent Development Kit (ADK).
What it offers:
🛠️ Build with the Agent Development Kit (ADK)
🏆 $50K cash prize pool
🧑🏫 Insights from Google experts
Sign up link below 👇
What it offers:
🛠️ Build with the Agent Development Kit (ADK)
🏆 $50K cash prize pool
🧑🏫 Insights from Google experts
Sign up link below 👇
May 14, 2025 at 6:00 AM
Google is running an Agents hackathon using the Agent Development Kit (ADK).
What it offers:
🛠️ Build with the Agent Development Kit (ADK)
🏆 $50K cash prize pool
🧑🏫 Insights from Google experts
Sign up link below 👇
What it offers:
🛠️ Build with the Agent Development Kit (ADK)
🏆 $50K cash prize pool
🧑🏫 Insights from Google experts
Sign up link below 👇
We just dropped Gemini 2.0 Flash Image Generation. Its available on Vertex AI (preview) and AI Studio.
Offering better visual quality and more accurate text rendering.
Notebook below to get you started 👇
Offering better visual quality and more accurate text rendering.
Notebook below to get you started 👇
May 8, 2025 at 3:39 AM
We just dropped Gemini 2.0 Flash Image Generation. Its available on Vertex AI (preview) and AI Studio.
Offering better visual quality and more accurate text rendering.
Notebook below to get you started 👇
Offering better visual quality and more accurate text rendering.
Notebook below to get you started 👇
Gemini 2.5 Pro just got an upgrade.
It' got better at coding, with significant gains in front-end web development, editing, and transformation. We also fixed a bunch of function calling issue.
You can access it as gemini-2.5-pro-preview-05-06 on VertexAI and AIStudio
More details in 🧵
It' got better at coding, with significant gains in front-end web development, editing, and transformation. We also fixed a bunch of function calling issue.
You can access it as gemini-2.5-pro-preview-05-06 on VertexAI and AIStudio
More details in 🧵
May 7, 2025 at 4:06 AM
Gemini 2.5 Pro just got an upgrade.
It' got better at coding, with significant gains in front-end web development, editing, and transformation. We also fixed a bunch of function calling issue.
You can access it as gemini-2.5-pro-preview-05-06 on VertexAI and AIStudio
More details in 🧵
It' got better at coding, with significant gains in front-end web development, editing, and transformation. We also fixed a bunch of function calling issue.
You can access it as gemini-2.5-pro-preview-05-06 on VertexAI and AIStudio
More details in 🧵
This is cool. You can now use Gradio to build a MCP server in just a few lines of Python.
May 6, 2025 at 10:12 AM
This is cool. You can now use Gradio to build a MCP server in just a few lines of Python.
Google just published a 2nd 76-page whitepaper on AI Agents that covers more advanced topics:
> Agentic RAG
> Agent Evaluation
> Multi-Agent Evaluation
> Production Architectures
If you're comfortable with agent basics and looking to understand the next layer this is a great read 👇
> Agentic RAG
> Agent Evaluation
> Multi-Agent Evaluation
> Production Architectures
If you're comfortable with agent basics and looking to understand the next layer this is a great read 👇
May 6, 2025 at 3:31 AM
Google just published a 2nd 76-page whitepaper on AI Agents that covers more advanced topics:
> Agentic RAG
> Agent Evaluation
> Multi-Agent Evaluation
> Production Architectures
If you're comfortable with agent basics and looking to understand the next layer this is a great read 👇
> Agentic RAG
> Agent Evaluation
> Multi-Agent Evaluation
> Production Architectures
If you're comfortable with agent basics and looking to understand the next layer this is a great read 👇
Interesting blog post by
@weightsbiases.bsky.social on the Agent2Agent (A2A) protocol.
The blog talks talks about how it works, its core components and how to use it with CrewAI and @langchain.bsky.social.
Blog below 👇
@weightsbiases.bsky.social on the Agent2Agent (A2A) protocol.
The blog talks talks about how it works, its core components and how to use it with CrewAI and @langchain.bsky.social.
Blog below 👇
May 5, 2025 at 1:22 AM
Interesting blog post by
@weightsbiases.bsky.social on the Agent2Agent (A2A) protocol.
The blog talks talks about how it works, its core components and how to use it with CrewAI and @langchain.bsky.social.
Blog below 👇
@weightsbiases.bsky.social on the Agent2Agent (A2A) protocol.
The blog talks talks about how it works, its core components and how to use it with CrewAI and @langchain.bsky.social.
Blog below 👇
Gemma 3 fine-tuning just got a serious boost! 🚀 Axolotl v0.8.0 delivers:
- Up to 3.6x speedup vs HF+FA2
- ~50% memory reduction
- Now includes Sequence Parallelism, Cut-Cross Entropy & Liger kernels
Plus you can fine-tune Gemma-3 27B with just 21.6GB VRAM!
- Up to 3.6x speedup vs HF+FA2
- ~50% memory reduction
- Now includes Sequence Parallelism, Cut-Cross Entropy & Liger kernels
Plus you can fine-tune Gemma-3 27B with just 21.6GB VRAM!
April 30, 2025 at 4:51 AM
Gemma 3 fine-tuning just got a serious boost! 🚀 Axolotl v0.8.0 delivers:
- Up to 3.6x speedup vs HF+FA2
- ~50% memory reduction
- Now includes Sequence Parallelism, Cut-Cross Entropy & Liger kernels
Plus you can fine-tune Gemma-3 27B with just 21.6GB VRAM!
- Up to 3.6x speedup vs HF+FA2
- ~50% memory reduction
- Now includes Sequence Parallelism, Cut-Cross Entropy & Liger kernels
Plus you can fine-tune Gemma-3 27B with just 21.6GB VRAM!
Qwen 3 is out. Some highlights:
↔️Offers a wide range of sizes, from 0.6B to 235B.
🧠 Includes MoE variants.
🎛️ Hybrid Thinking Control with `enable_thinking=True`.
🌍 Multilingual Support: with support for 119 languages and dialects.
↔️Offers a wide range of sizes, from 0.6B to 235B.
🧠 Includes MoE variants.
🎛️ Hybrid Thinking Control with `enable_thinking=True`.
🌍 Multilingual Support: with support for 119 languages and dialects.
April 29, 2025 at 2:58 AM
Qwen 3 is out. Some highlights:
↔️Offers a wide range of sizes, from 0.6B to 235B.
🧠 Includes MoE variants.
🎛️ Hybrid Thinking Control with `enable_thinking=True`.
🌍 Multilingual Support: with support for 119 languages and dialects.
↔️Offers a wide range of sizes, from 0.6B to 235B.
🧠 Includes MoE variants.
🎛️ Hybrid Thinking Control with `enable_thinking=True`.
🌍 Multilingual Support: with support for 119 languages and dialects.
We've just launched Gemini 2.5 Flash.
⚡️High quality for best cost
🧩 Ability to dynamically think on complex tasks
🎛️ Configurable thinking budget (decide how many tokens to allocate to "thinking".
⚡️High quality for best cost
🧩 Ability to dynamically think on complex tasks
🎛️ Configurable thinking budget (decide how many tokens to allocate to "thinking".
April 18, 2025 at 12:06 AM
We've just launched Gemini 2.5 Flash.
⚡️High quality for best cost
🧩 Ability to dynamically think on complex tasks
🎛️ Configurable thinking budget (decide how many tokens to allocate to "thinking".
⚡️High quality for best cost
🧩 Ability to dynamically think on complex tasks
🎛️ Configurable thinking budget (decide how many tokens to allocate to "thinking".
Setting up Gemma 3 with Ollama on Google Colab? I wrote down the steps I followed, including enabling streaming responses and handling multimodal inputs.
Details & steps on the blog ⤵️
Details & steps on the blog ⤵️
April 16, 2025 at 5:22 AM
Setting up Gemma 3 with Ollama on Google Colab? I wrote down the steps I followed, including enabling streaming responses and handling multimodal inputs.
Details & steps on the blog ⤵️
Details & steps on the blog ⤵️
Google Cloud Next 2025 and the ADK launch done ✅
On to the next.
On to the next.
April 12, 2025 at 1:02 AM
Google Cloud Next 2025 and the ADK launch done ✅
On to the next.
On to the next.
Quickstart:
google.github.io/adk-docs/get...
google.github.io/adk-docs/get...
April 9, 2025 at 2:41 PM
Quickstart:
google.github.io/adk-docs/get...
google.github.io/adk-docs/get...
Get hands-on with Gemma 3 without the setup headache?
Hereby an updated guide that helps you get started with Gemma 3 and
@ollama
on Colab for a fast start. Based on one of our most popular 2024 posts. ⤵️
Hereby an updated guide that helps you get started with Gemma 3 and
@ollama
on Colab for a fast start. Based on one of our most popular 2024 posts. ⤵️
April 3, 2025 at 5:57 AM
Get hands-on with Gemma 3 without the setup headache?
Hereby an updated guide that helps you get started with Gemma 3 and
@ollama
on Colab for a fast start. Based on one of our most popular 2024 posts. ⤵️
Hereby an updated guide that helps you get started with Gemma 3 and
@ollama
on Colab for a fast start. Based on one of our most popular 2024 posts. ⤵️
You can now deploy and use Gemma 3 in only three lines of code.
Code below⤵️
Code below⤵️
April 1, 2025 at 4:27 AM
You can now deploy and use Gemma 3 in only three lines of code.
Code below⤵️
Code below⤵️
Exciting news! Gemini 2.5 is now also available on Vertex AI. Plus you can try it out in Express Mode. No GCP login required!
You can find it using the model ID gemini-2.5-pro-exp-03-25 through both the UI and SDK.
See below how to get started ⤵️
You can find it using the model ID gemini-2.5-pro-exp-03-25 through both the UI and SDK.
See below how to get started ⤵️
March 29, 2025 at 1:52 AM
Exciting news! Gemini 2.5 is now also available on Vertex AI. Plus you can try it out in Express Mode. No GCP login required!
You can find it using the model ID gemini-2.5-pro-exp-03-25 through both the UI and SDK.
See below how to get started ⤵️
You can find it using the model ID gemini-2.5-pro-exp-03-25 through both the UI and SDK.
See below how to get started ⤵️
The Gemma 3 technical report is now available on arXiv with some great insights in the improvements and development process. ⤵️
arxiv.org/abs/2503.19786
arxiv.org/abs/2503.19786
March 28, 2025 at 3:43 AM
The Gemma 3 technical report is now available on arXiv with some great insights in the improvements and development process. ⤵️
arxiv.org/abs/2503.19786
arxiv.org/abs/2503.19786
Diving into Multimodal Fine-Tuning with Gemini 2.0. This opens doors for more fine-tuned AI agents capable of richer reasoning.
Notebooks to get you started below ⤵️
Notebooks to get you started below ⤵️
March 27, 2025 at 6:23 AM
Diving into Multimodal Fine-Tuning with Gemini 2.0. This opens doors for more fine-tuned AI agents capable of richer reasoning.
Notebooks to get you started below ⤵️
Notebooks to get you started below ⤵️
We just released Gemini 2.5 Pro, our most intelligent model with impressive capabilities:
🧮Enhanced math, science, and reasoning
🖥️Significantly improved coding capabilities
📏1 million context window
It's available on Google AI Studio now, with Vertex AI coming soon. Code below to get you started ⤵️
🧮Enhanced math, science, and reasoning
🖥️Significantly improved coding capabilities
📏1 million context window
It's available on Google AI Studio now, with Vertex AI coming soon. Code below to get you started ⤵️
March 26, 2025 at 1:27 AM
We just released Gemini 2.5 Pro, our most intelligent model with impressive capabilities:
🧮Enhanced math, science, and reasoning
🖥️Significantly improved coding capabilities
📏1 million context window
It's available on Google AI Studio now, with Vertex AI coming soon. Code below to get you started ⤵️
🧮Enhanced math, science, and reasoning
🖥️Significantly improved coding capabilities
📏1 million context window
It's available on Google AI Studio now, with Vertex AI coming soon. Code below to get you started ⤵️
Excited to see the that we now have a lightweight Gemma library. Just run `pip install gemma` and you are good to go.
Below you'll find docs that include fine-tuning, sharding, LoRA, PEFT, multimodality and tokenization in LLMs.
Below you'll find docs that include fine-tuning, sharding, LoRA, PEFT, multimodality and tokenization in LLMs.
March 24, 2025 at 8:41 AM
Excited to see the that we now have a lightweight Gemma library. Just run `pip install gemma` and you are good to go.
Below you'll find docs that include fine-tuning, sharding, LoRA, PEFT, multimodality and tokenization in LLMs.
Below you'll find docs that include fine-tuning, sharding, LoRA, PEFT, multimodality and tokenization in LLMs.
Obsessed with LLM Leaderboards? You're not alone. But leaderboards don't guarantee a model fits your needs. Build a solid evaluation strategy with key metrics to de-risk your model selection.
The guide below is a great starting point to learn more about evaluation. ⤵️
The guide below is a great starting point to learn more about evaluation. ⤵️
March 21, 2025 at 4:59 AM
Obsessed with LLM Leaderboards? You're not alone. But leaderboards don't guarantee a model fits your needs. Build a solid evaluation strategy with key metrics to de-risk your model selection.
The guide below is a great starting point to learn more about evaluation. ⤵️
The guide below is a great starting point to learn more about evaluation. ⤵️