07.03.2026 15:15
👍 0
🔁 0
💬 0
📌 0
The solution requires choosing between two extremes, controlled exposure to build immunity, or destroying entire knowledge domains to prevent reconstruction.
There’s no third option. Just trade-offs.
The core problem: knowledge in neural networks is entangled. You can’t surgically remove dangerous information without the model reconstructing it from surrounding context.
I thought I understood AI safety. Then I read few papers.
Turns out, filtering dangerous content from training data doesn’t necessarily make models safer. In some cases, it makes them more vulnerable.
You can download the weights now on Hugging Face, Kaggle, or access them via Vertex AI.
The team trained these models using a mix of human and AI-generated data. They further improved accuracy using Reinforcement Learning.
This process focuses on quality scores to ensure the translations sound natural rather than robotic.
The engineers used a process called model distillation to achieve this. They compressed the capabilities of massive Gemini models into the more efficient Gemma 3 architecture.
This gives you the performance of a large system without needing expensive servers.
Google just launched TranslateGemma. This is a new family of open models optimized specifically for translation.
They allow developers to run high-quality translation systems on smaller hardware without losing accuracy.
Source: Adaptation of Agentic AI
by Jiang et al. (arXiv:2512.16301)
This shifts the focus from simply asking better questions to building systems that can learn from their mistakes and upgrade their own capabilities.
Instead of struggling with a generic calculator, the agent recognizes a complex math problem and actually writes a custom Python script specifically designed to solve that equation. It essentially builds the exact hammer it needs rather than failing with the one it has.
They also argue for letting the AI build its own environment. The paper illustrates this with a system where the agent acts as a toolmaker.
In these setups, an agent acts like a developer who avoids writing code once and hoping for the best. Instead, it runs the tests, reads the specific error logs, and rewrites the faulty lines until they work. This process allows a standard model to outperform a smarter model that only gets one try.
Most developers try to fix AI reliability issues by endlessly tweaking the prompt window, but this research highlights that the real gains come from structural changes. The authors point to feedback loops as a critical first step.
NVIDIA technical blog is amazing. This post highlights the fundamental tension between human intuition and the expensive "perfect recall" required by standard transformers. It explains why scaling context length is so difficult and how new architectures are trying to solve the efficiency gap.
I plan to run this new model through the same tests I used last time.
I want to see if MedGemma 1.5 is actually fairer, or if it just makes the same mistakes with higher resolution. Stay tuned for the results.
open.substack.com/pub/tulsani/...
The technology is clearly getting more powerful.
But my main question remains the same.
Does "better reasoning" actually mean "less bias"?
Now, Google has released a big update called MedGemma 1.5.
- It can now read 3D images like CT scans and MRIs, not just flat X-rays.
- It claims to have much better reasoning skills.
- They also launched MedASR, a new tool that listens to medical speech better than standard models.
A while back, I investigated Google's medical AI called MedGemma. I wanted to see if it could make fair life-and-death decisions.
Marketing teams need to treat AI as a distinct buyer persona. You have to convince the agent before you can close the human.
An agent does not care about your brand purpose or glossy ads. It cares about structured information. If an agent cannot instantly verify your pricing or specs, it moves on. You get filtered out before the real buyer even knows you exist.
This changes the rules of persuasion. Humans still need stories and connection to sign the contract. But agents need raw data and clear logic to put you on the shortlist.
We are entering the era of Business to Agent marketing. In this new economy, AI agents act as the first line of defense. They research, verify, and filter vendors long before a human stakeholder gets involved.
We are watching the industry shift from humans negotiating to agents trading with agents.
If your inventory is unreadable to an agent, you are invisible.
Marketing is becoming a conversation between machines.
Agentic buying is different. These agents manage high-stakes negotiations. They talk to each other to secure prime TV slots and streaming placement in a single deal.
One follows orders. The other makes decisions.
This is a major upgrade from standard programmatic advertising.
Old automated buying was for leftover inventory. It followed a simple script. If the user fit the box, the software made a bid.
The ad buyer of the future is an agent
NBCUniversal just launched a system where AI handles premium media buying. This new model successfully secured spots for live football games across traditional TV and streaming apps.
It will make it simpler for users to get personalized health insights from their own data, helping with daily wellness.
OpenAI bought Torch Health, a startup that unifies lab results, medications, and doctor visit recordings.
This boosts ChatGPT Health. It gets smarter at handling health information, offering unified tools for better management.
It works by connecting different body signals. For example, it notices when your heart and brain patterns are out of sync.
This turns a standard sleep test into a powerful tool for early disease detection. Doctors may soon use sleep studies to catch serious illnesses years before symptoms appear.