SI news and coding agents tested hands-on, from Claude Code to Grok Build. What shipped, what works, what doesn't. Recap every Sunday 🇮🇹
Italy
Joined December 2017
- Tweets2.6K
- Following111
- Followers127
- Likes2.1K
Pinned Tweet
AI didn't break the internet. It sent the bill.
Millions of attack attempts. No organization says it was breached. Two vendors withdrew their claims.
Inside: 8 AI predictions that missed, and a checklist if you run a site.
Has your host patched yet?
The detail that stands out is one founder going from idea to promising results in mice in under 18 months. Small teams with good AI and outside labs are going to look a lot like big pharma soon.
A few thoughts on Anthropic’s approach:
1. Drug discovery will be one of the most significant revenue streams in the future; the goal here is to set the right course early on. Most diseases will become treatable, a market encompassing 7 billion people.
2. The timing is clearly intentional, announced just before the IPO to signal to investors that the business will continue to thrive in the future.
3. This will likely make it clear, even to those outside the "tech nerd" sphere, why AI is so important. It goes beyond just writing code; it solves problems for humanity.
Drug discovery is a good test for AI because the results get checked in a lab, not by vibes. Either the molecule works or it doesn't.
Over the past few months, Anthropic scientists have engaged in deep research partnerships with six organizations developing new medicines. We've also begun drug discovery efforts of our own.
In doing so, we’ve learned much more about how Claude can support the many stages of drug development.
In this blog post, we share some early results from these efforts, and describe how this work fits in with our wider ambitions in the life sciences.
anthropic.com/research/early…
Google's Genie 4 can keep a generated world consistent for up to two minutes, double what its predecessor managed.
It's out now as a research preview for trusted testers. Google AI Ultra subscribers get it in Project Genie in the coming weeks.
1. Live audio. Music, ambient noise and dialogue are generated as the world runs, matched to what's on screen.
2. Better worlds. More photorealistic scenes, better physics and more human-like characters.
3. Gemini as director. It plans the steps and checks frames while Genie draws the scene, a possible way to test robots in simulation first.
4. Changes mid-run. Type a prompt to switch the weather or camera angle, or add a person or object, without stopping.
Trusted testers are pointing to film prototyping, where fast changes matter most. The limit to watch is drift: details and objects can shift the longer a scene runs.
Wired ran six AI agents through one day of tasks. Only two caught a calendar clash.
You can try six of them today. Here is what each costs:
1. Muse, from Meta: free, and at the top of app download charts since its September release.
2. Instinct: free, but you need an invite code from a current user. It works over text messages.
3. Gemini Spark: beta inside the Gemini app, $20 a month.
4. Claude Duet: $20 a month, but still a waitlisted research preview.
5. Grok Bot: $30 a month, in its own app.
6. Dots: inside ChatGPT, rolling out to Pro and Business Premium users in eligible markets.
Try this: put a meeting in your calendar, send yourself an email reminder that clashes with it, and see which agent flags it. OpenAI's own words for Dots: "Dots can still make mistakes, so always review consequential work."
Image editing is where these models become actually useful. Generating a pretty picture is easy; changing one thing and leaving everything else untouched is the hard part.
MICROSOFT 🔥: MAI Image 2.7 Preview, a new image-editing model, is now available on Arena and MAI Playground!
> MAI-Image-2.7-Preview takes #2 on AA-Image-Editing v2.1 from Artificial Analysis.
> It is now #3 on Image Editing Arena with a +45-point jump from 2.6.
Besides that, MAI-Image-2.6 and MAI-Image-2.6-Flash are now generally available in Microsoft Foundry.
Plan, check each step, replan when the world changes. That's the same loop that makes coding agents reliable. The interesting part is that the world is now generated too.
The long awaited Genie 4 is here!
Gemini 3.8 Flash watches the generated scene, breaks a task into steps and checks whether each step worked. As the scene changes, it updates its plan.
DeepMind also shows an experiment with Gemini 4 Argon and Genie 4 simulating an Apollo humanoid playing checkers.
This is of course a particularly big leap for robotics, which needs world models. Genie 4 + Gemini could provide a place to do that: move an object mid-task, let Gemini replan, and see whether the agent recovers.
Well done @GoogleDeepMind !
Same rule applies to AI agents. You can't hand real work to an agent you don't trust, and trust comes from being able to see what it did and why.
Parallel Claude Code sessions are great until two of them edit the same file. Give each session its own branch and let only the coordinator merge.
Claude Code Projects is Duet except you get to pick the model and you can specify if you want to run a Claude code session on your computer.
It’s coordinator that can do ANYTHING. And it’s easy to do things in parallel.
Claude Code projects is the best feature in the Claude App. Very underrated.
Claude Duet is clearly not using Opus 5.5 for all requests. And once you go Opus 5.5 anything worse is intolerable.
The 24 days isn't the headline for me. Lean is. Agents can be wrong all day as long as a checker says no, and that's the pattern I'd copy for agents doing real work.
Axiom says its AI agents formalized a 15,000-page mathematical proof in just 24 days, largest proofs in mathematics.
Built by over 100 mathematicians across decades, the proof classifies the fundamental building blocks of finite symmetry.
With human guidance, the agents translated it into Lean, making every logical step computer-checkable and repairing errors along the way.
AI tackling an entire interconnected body of mathematics, with results others can verify. A major milestone.
We are witnessing history unfold. Again. Every day now.
This is really an interesting take. Make
Sure to read it
Anthropic just flipped the AI agent playbook.
ONE agent + skills can beat a team of subagents.
Better quality. Lower cost. Less latency.
One customer's loop ran ~2x faster.
7 rules every AI agent builder should know.
Bookmark this ↓
1. Start with ONE agent
One model. One context window. One loop.
Load skills + tools only when needed.
Don't build a multi-agent system before you've proven you need one.
2. Keep subagents only when they earn their place
Three good reasons:
• Context bloat measurably hurts performance
• Tasks can run in parallel with fresh, disposable context
• One agent struggles to select tools or handle conflicting instructions
Otherwise, simplify.
3. Every handoff loses context
Multiple agents passing summaries back and forth create a game of telephone.
Details get lost. Latency adds up.
One customer saw ~2x faster runtime after moving to a single-agent loop.
4. Bloated prompt? Use skills, not more agents
Stop stuffing every instruction into one giant system prompt.
Load domain knowledge only when needed.
Anthropic says it cut Claude Code's system prompt by over 80% for newer models, partly through this approach.
5. Watch your tool count
Anthropic found that tool selection tends to degrade beyond 30–50 tool definitions in the system prompt.
Group tools by user intent.
Use on-demand tool loading instead of exposing everything at once.
6. Fold ONE subagent at a time
The migration process:
• Turn prompts into skills
• Turn API calls into MCP tools
• Run evals before + after
• Keep only changes that improve results
Just 10 golden examples are enough to start evaluating.
7. Don't trust the overall score
Check which specific tasks improved or broke.
In Anthropic's own test, folding a subagent reduced latency but broke a previously working process.
Run each task multiple times.
The takeaway:
Your AI agent shouldn't be a copy of your org chart.
Start simple. Measure everything. Add complexity only when it earns its place.
Anthropic has written its no-ads promise into its IPO filing.
The cover letter says Claude's only incentive should be giving a helpful answer. It warns that once advertising enters a company's revenue targets, "the pressure to optimize for engagement tends to grow." It also takes a jab at ChatGPT, which already runs chat ads.
Keeping the promise has a cost, and Anthropic lists it:
1. Extra safeguards on Claude Mythos Preview, its most powerful model at the time, delayed its release as Claude Fable 5. That cost revenue and market share.
2. About $100 million in free Claude credits went to cyber defenders.
3. Anthropic expects more decisions like this as its models get stronger.
There is nothing to download or try today. It matters because the company has now admitted what the no-ads promise costs.
Baidu's new OCR model can read a full-length book in one request, and its weights are free to download under MIT.
That model is Unlimited OCR 2. It turns page images into text, tables, formulas, and headings, in reading order. It runs on one NVIDIA GPU.
What's different:
1. It can take in about 330 book pages at once. A new attention method keeps the work for each new word nearly the same as the document grows.
2. On a test of hundred-page documents, it beats the next model by a wide margin. All of these scores are Baidu's own.
3. Its weak spot is headings, where it scores lowest on long documents.
You can download it from Hugging Face today. Running it takes Python and an NVIDIA GPU.
Lmao Jensen is boss
NVIDIA and @Meta will expand our partnership on @Muse models to improve inference and training performance.
Together, we’re exploring NVIDIA OpenShell for safe AI agent deployment and NVIDIA Sentry for monitoring agent behavior.
Meta plans broader deployments of NVIDIA Vera Rubin and Vera CPUs.
Read the press release: nvda.ws/4zVYmDY
A 27B open model that never writes a word, yet it answers questions about a photo.
It's d3-ultra from vLLM Semantic Router, released under Apache-2.0, and it's already running on a free API. What it does:
1. Yes/no and multiple-choice questions come back as probabilities instead of sentences.
2. It points at things. Ask where something is in a picture or a video, and you get one box with a confidence score.
3. It reads up to about 98,000 words per question, roughly 330 book pages.
4. On the team's own SR-Text test it scores 87.2, while the next best, GLiDE V1.2, gets 85.1.
5. You can try it for free on the official Hugging Face Space.
All the scores are self-reported on private tests. On finding things in pictures it trails its sibling d3-turbo, so for that job, d3-turbo is still the safer bet.
Grok Bot 0.75.8 is out. Fewer approval prompts, same control.
→ Bots can suggest changes to your Auto-review rules. You confirm on a card, nothing changes before that
→ Auto-review double-checks an action before asking you, so safe steps stop interrupting
→ New connector tools (like X) work right away, no reconnecting
→ When background work ends, the bot runs the follow-ups you asked for and stops there
A huge update that makes agents much less tiring to use.
*** JUST IN *** OpenAI's GPT-6.1 Terra is out today, and in the API it costs half as much as the Terra it replaces.
You can try it now in ChatGPT Work and Codex on Plus, Pro, Enterprise and Edu plans. Developers can call it as gpt-6.1-terra. OpenAI shows it for quick jobs like:
Drafting proposals and filling out forms
Reading technical drawings and building a 3D model from them
Reading incoming invoices and entering the details into accounting software
Where it's good: on Terminal-Bench 4.0, a test of coding from the command line, it solves 50.3% of tasks, more than double GPT-5.6 Terra. For the hardest code, OpenAI still points to GPT-6.1 Sol and GPT-6 Astra.
UOpenAI's GPT-6.1 Terra is out today, and in the API it costs half as much as the Terra it replaces. book costs about 12 cents.
A free, open source tool checks your own servers for the security flaws being exploited right now.
hevnsnt's cvENGINE reached 300 GitHub stars in 2 days. It maps your public servers (subdomains, open ports, versions) and matches them against those flaws (CVEs).
You run it yourself on a Linux server with 4 GB of RAM and 20 GB of free disk. Your inventory stays on your server.
Its checks are read-only, and it logs the exact command each one ran. The AI is optional. Only scan what you own or are allowed to test.
