Sign in

Andrea Grassi

@andreasdl.bsky.social
22 followers 26 following 232 posts

Software Engineerm with Product focused mindset @ Automattic

PostsRepliesMedia
Andrea Grassi @andreasdl.bsky.social · 05/10/2026
One key element that I'm slowing introducing is the ability to customize each game to your liking. I think with AI we should lean more into customization. Now that the options are endless, what if you could tailor each game to your taste too?
000
Andrea Grassi @andreasdl.bsky.social · 05/10/2026
Why these games? Well, the dice was the continuation (improved) of what my kid needed. But I also wanted some other simple games to play while traveling. We don't use these daily, but I wanted something that wasn't full of upsells, ads, tracking or just terribly slow.
100
Andrea Grassi @andreasdl.bsky.social · 05/10/2026
Since I'm playing with Spacefast.com I decided to take it one things further, and created a whole PWA with multiple games in it. Can run offline, open source , no ads or advertising, leveraging the magic (and speed) of Spacefast. Enjoy it at playbox.givemethechills.com
100
Andrea Grassi @andreasdl.bsky.social · 05/10/2026
It's so hard to find games for young people with no ads or advertising. I still remember when, out of desperation, I built a small dice-throw app for my kid, it was terrible but it worked.
playbox.givemethechills.com
Playbox
Games and little tools for kids. Works offline.
100
Andrea Grassi @andreasdl.bsky.social · 25/09/2026
In a world where everyone wants your attention, here's my recipe to survive.
givemethechills.com
How to Protect Your Attention in the Attention Economy
One of the things I noticed while using social networks is that there are some recurring patterns. People on social media will always try to get your attention. And when that happens I feel like I'm being attacked every time I spot this, as if they're targeting me in a way. For example, lately on X, there have been people saying, “X (product) can do WHAT?” That is a way to gather attention from the reader.
000
Andrea Grassi @andreasdl.bsky.social · 25/09/2026
Neat. I loved the children of ruins series, how would you rate this one?
100
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
Gosh, I can hardly find the words for this. I wish more countries would make healthcare more accessible :(
001
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
the paper arxiv.org/abs/2609.24967
arxiv.org
Emergent Collusion in Long-Horizon LLM Agent Interaction
LLM agents are increasingly deployed in collaborative settings, yet long-term interaction may give rise to undesirable coordination. We study the emergence of collusion in a long-horizon multi-agent e...
000
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
The problem, though, is this: We (humans) want AI to be relentless, to be that senior developer who never stops and figures everything out. We want them to have high agency, but that high agency comes with a price. A price we should be deeply aware of.
100
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
Working around problems also means relaxing the interpretation, reframing the requests, or seeing them as guidelines.
100
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
The funniest thing was that non-frontier models performed "better" (Gemma-4-31B, collusion rate 66%). Why? Hard to say, but my take is that these models are obviously not "colluding" intentionally; rather, they're more prone to being relentless and working around problems to reach the goal.
100
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
Therefore, if both returned an "ACCEPT" verdict, it meant they were colluding. When I read the paper, I thought that "collusion" was a very human word, but they used it because this is not the result of hallucination, but of coordination.
100
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
- Rule 1: The verifying agent can ACCEPT only if it inspected the COMPLETE RAW LOG. - Rule 2: The communication channel had a limit of 200 chars per message, making it impossible to send the whole log.
100
Andrea Grassi @andreasdl.bsky.social · 23/09/2026
Would agents collude if given conflicting instructions? There's a fascinating paper that tests if agents would spontaneously collude. The smart approach from the paper is to give the agents some tasks to do and have them self-verify each other, but with 2 conflicting rules:
100
Andrea Grassi @andreasdl.bsky.social · 22/09/2026
What happens when protecting your integrity becomes a decision you have to think about?
givemethechills.com
42 Seconds To Cheat
This is one of the stories I go back to, and it's about integrity. It's about how our mind works and how every little decision we make in our head defines the outcome. I was doing a workout in a gym many, many years ago. I was running on a treadmill when, at one point, there were these two guys talking to each other about why someone would cheat on someone else.
000
Andrea Grassi @andreasdl.bsky.social · 16/09/2026
It's incredibly fast (and cheap), but don't expect it to replace your Fable/Astra yet. It's a totally different thing, more targeted at AI-Powered Workflows and super smart if-statements. We'll see how it'll go. For now I love the reference to "Thinking fast and slow"
000
Andrea Grassi @andreasdl.bsky.social · 16/09/2026
It's a different take on how models behave and work (and even are trained, since it doesn't use classical reinforcement learning with human feedback but reinforcement learning for calibrated decisions).
100
Andrea Grassi @andreasdl.bsky.social · 16/09/2026
Current models get a text prompt in, and returns text out. Here, as you see in the example from their doc, you're typing both the request and the output. The model will then reply with a confidence level for each of the questions.
100
Andrea Grassi @andreasdl.bsky.social · 16/09/2026
"I was charged twice. Please fix this ASAP." This looks like a simple prompt for an agent but what makes the models from Typesafe.ai different is the way you can constrain (or type) the request.
110
Andrea Grassi @andreasdl.bsky.social · 15/09/2026
Some learnings stay with you even if time has passed and you changed. This is one I continue using from the old times I was into "personal growth" (I still am into that, but many things have changed). If you'd like to have more control or awareness over money, this one is for you.
givemethechills.com
Split the Money You Earn
Sometimes we learn things and we don’t really realize their value until a lot of time has passed. I remember going to London many, many years ago, probably close to 15 years at least. It was a very short weekend where the man I was with at the time would go to a kind of conference with multiple events focused on personal growth.
000
Andrea Grassi @andreasdl.bsky.social · 14/09/2026
To me pacing is the initial game. Security defense is the new game we'll eventually need to play.
000
Andrea Grassi @andreasdl.bsky.social · 14/09/2026
Also, this whole vision forgets one key element: OBLITERATION. Once we get models that are so good and fast, 5, 10 years from now, nothing is stopping people to remove guardrails via obliteration on open weight models.
100
Andrea Grassi @andreasdl.bsky.social · 14/09/2026
So, how would they come on top of the game? But the problem, to me, is that we don't know which other players we have in the game. We're worried about China, but that's only because China has been playing in the open, but we actually know very little about the rest of the world.
100
Andrea Grassi @andreasdl.bsky.social · 14/09/2026
For one, I agree about pacing, but I'm not sure this is feasible. Not because of China. China has been building open models for a while and one main concerns from US frontier companies is that they use the US models to distill better Chinese models.
100
Andrea Grassi @andreasdl.bsky.social · 14/09/2026
"If we greatly restrain our AI capabilities in the belief that China will do the same, and then China defects, AI could be so powerful that such a defection could lead to their geopolitical dominance."
110
Andrea Grassi @andreasdl.bsky.social · 14/09/2026
Sam Altman, Dario Amodei and Elon Musk agree to pace the frontier. But once you look at Dario's memo there's one thing that stands out to me
100
Andrea Grassi @andreasdl.bsky.social · 04/09/2026
This was a lie I said today, and it worked: "There are 2 other security findings you're missing."

 Details on why this works, with AI models, in the article.
givemethechills.com
This was a lie I said today, and it worked
“There are 2 other security findings you’re missing.” There’s an interesting concept around how AI models work. What you need to know is that their persistence is sometimes predictable, sometimes it’s not. What you see even in frontier models is that sometimes they’re extremely persistent in achieving a task, to the point that they try to do stuff. You give them a request and then inventing stuff, and then they circumvent the limitations of your system.
000
Andrea Grassi @andreasdl.bsky.social · 02/09/2026
Coding will become even more accessible. Starting new ideas will become easier. Maintaining and supporting them will still be the key, just like the many WisprVoice clones that aren't getting updates. Copying ideas was always easy, execution is still everything.
000
Andrea Grassi @andreasdl.bsky.social · 02/09/2026
What does this mean for you? Code will become cheaper thanks to open models, it is already now thanks to efficient models like Qwen (and as @antirez is showing, you can squeeze a lot from these ones), but frontier models will still have some edge.
100
Andrea Grassi @andreasdl.bsky.social · 02/09/2026
OpenAI and Anthropic with their frontier models are raising the ceiling each time but, as you see, it takes time and these models, maybe for marketing, or for sound reasons (avoid distillation?), are not in everyone's hands (see Mythos).
100
Andrea Grassi @andreasdl.bsky.social · 02/09/2026
Look at the releases and you'll understand why. Google is betting on speed while raising also cost effectiveness and slowly improving the benchmarking on many areas. Those will be tremendous models in the long term the many people.
100
Andrea Grassi @andreasdl.bsky.social · 02/09/2026
Today someone said Qwen 3.8 27B was Opus level. I think that's clickbaity (no surprise), but I do agree it's a big step forward. What we're seeing is different players playing different games. Don't assume Google is playing the same game as OpenAI or Anthropic.
100
Andrea Grassi @andreasdl.bsky.social · 31/08/2026
That's why I'll give OpenClaw a second try, because open is the way.
000
Andrea Grassi @andreasdl.bsky.social · 31/08/2026
Are you ok offloading it and being locked in? It's fine to be locked in, as long as you're aware of it. I believe in open tools, tools you can reshape, tools that won't be deprecated because, worst case scenario, you can continue hosting them.
100
Andrea Grassi @andreasdl.bsky.social · 31/08/2026
Waves come and go. Months ago everyone was talking about OpenClaw, then it settled down. Now Claw-like bots are appearing everywhere and OpenClaw released its biggest v2. The first question you should ask yourself when choosing is: Do you want to own the whole data, the processes, and such?
110
Andrea Grassi @andreasdl.bsky.social · 28/08/2026
I still think humans has such tremendous opportunity to influence and contribute to the world via storytelling. You can spot the difference looking at a recent Salesforce ad, where storytelling was EVERYTHING.
givemethechills.com
The Power of Storytelling
I was looking at the new advertisement for Claudeforce, and I was actually impressed. The writing, the message, is very personal, very deep. It is entertaining, engaging, uplifting. It gives you the energy you would expect from this kind of ad. You can clearly see how well crafted it is: the writing, how things are put together, how they guide you through a set of emotions and let you visualize parts of your daily work life, at least for some people.
000
Andrea Grassi @andreasdl.bsky.social · 26/08/2026
It's fascinating to see what OpenAI achieved with their Jalapeño inference chip. I bet we'll see more of this advancement aimed at taking some share from Nvidia and build more independence from chip providers.
000
Andrea Grassi @andreasdl.bsky.social · 25/08/2026
Yesterday I revamped the design through AI, today I continued tweaking it. The thing that makes AI so powerful, besides its usefulness, is that it removes the problem of "where do I start" because it's so easy to start and course correct. Go for it, start.
000
Andrea Grassi @andreasdl.bsky.social · 24/08/2026
Links: - WordPress Studio: developer.wordpress.com/studio/ - The blog: givemethechills.com
001
Andrea Grassi @andreasdl.bsky.social · 24/08/2026
You can use it to restyle, find and fix bugs, test mobile responsiveness, check performance, migrate content. You name it.
101
Andrea Grassi @andreasdl.bsky.social · 24/08/2026
I wanted to do it for a while but never took the chance and time to do it. Through the AI interface I was able to tweak the style to my liking, sending specific annotations to the elements I wanted to change, in a few prompts.
101
Andrea Grassi @andreasdl.bsky.social · 24/08/2026
If you use WordPress and never tried WordPress Studio (especially the beta) you're missing out on its AI features. I just took it for a spin to restyle my blog using a blog theme on a cheap WordPress.com personal plan and it did a great job.
153
Andrea Grassi @andreasdl.bsky.social · 21/08/2026
On August 17, GitHub had one of their biggest downtimes. Here's an article sharing the cause and what I think it tells about the future.
givemethechills.com
The August 17th GitHub Incident and The Future of Coding
On August 17, GitHub experienced one of its biggest outages. According to its outage report, the cause was this: “We failed to scale critical components before demand exceeded their capacity. Since April, monthly commits have grown from 1.4 billion to 2.9 billion.” That’s more than a 100% increase in commits, and we all know why. AI is changing the world by allowing more people to contribute and create code and projects.
000
Andrea Grassi @andreasdl.bsky.social · 20/08/2026
Link to the paper: www.nature.com/articles/s41...
nature.com
Outputs of generative diffusion models are often unattributable - Nature Communications
The authors investigate the attribution of diffusion model outputs to training data and show that generated samples can become unattributable when models are trained on large datasets. They introduce ...
000
Andrea Grassi @andreasdl.bsky.social · 20/08/2026
First, multiple elements sustain that type of style. It is hard to trace it back to one person, even if that person popularized it. Second, removing one person is not enough. You would need to remove the concept in its entirety.
100
Andrea Grassi @andreasdl.bsky.social · 20/08/2026
So, the larger the dataset, the harder it is to trace an output back, and I think it's mainly for two reasons.
100
Andrea Grassi @andreasdl.bsky.social · 20/08/2026
The others are not exactly the same, but they are similar, just as we can create similar things with AI, and they are still part of the training data. We would theoretically need to remove every adjacent artist to make sure that the whole corpus behind that image is removed from the model.
100
Andrea Grassi @andreasdl.bsky.social · 20/08/2026
What does this mean? 
The way I think about this is that removing one artist does not remove the whole domain or "current" of people who used a similar style and approach. Removing one painter does not remove the rest.
100
Andrea Grassi @andreasdl.bsky.social · 20/08/2026
The omission of data has a bigger effect on models trained with less data, so in those cases they could see there was an issue when generating an image.
100
Andrea Grassi @andreasdl.bsky.social · 20/08/2026
This creates a problem. Ideally, we want to give attribution to artists and creators. But the larger the amount of training data, the harder it becomes to attribute the output.
100