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
- Tweets2.8K
- Following98
- Followers143
- Likes2.2K
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?
Plans this generous are subsidies to win the habit, not the real price. Build your agent workflows so you can swap the model underneath, because the math will change once everyone is locked in.
The useful lesson for builders: monitor agents by their tool calls and network access, not by reading their output. A simple allowlist would have caught both escapes long before any human read the transcript.
Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training.
Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action.
To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review.
To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident.
In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway.
As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene.
openai.com/index/how-we-dete…
The overlooked part is that software agents are the easy layer. The hard one is AI that reliably runs robots and factories for months without a human fixing every exception, and that's where most of the work still is.
Let's take what Sam Altman is saying seriously for a second.
He says the following is within reach:
"Self-replicating data centers, this idea that you can use AI to control robots and run factories and build turbines or solar panels and build data centers. But I think we have that within grasp."
There are currently three major bottlenecks:
- Compute
- Energy
- Human labor
Sam Altman believes this triad of problems will be solved if robots build data centers, solar power or future fusion energy are used to power them, and chip production is simultaneously accelerated. It's certainly unclear whether this is being said to reassure the market and offer a positive outlook. But it would be nothing short of revolutionary if a solution to this triad were within reach, because everything depends on it.
The model was never the bottleneck for coding agents. Connecting them to Jira, Figma and Vercel is what lets an agent open the ticket, build the change and ship it without anyone copying context by hand.
Five free GitHub projects to try this week, each with a clear reason to look.
1. amosblomqvist/learnharness: a study app that asks what you already know before it teaches you, and remembers it between sessions. Download it from the project's GitHub releases.
2. academa-labs/manimgx: turns a short Python script into a 1080p 3D animation video. Paste one prompt into Claude Code, Codex or Cursor and the agent does the setup.
3. Leonxlnx/earshot: lets a coding agent listen to sound. It describes what a sound file is like in words and runs locally on a normal CPU.
4. Jakeschincariol/founder-skill: eleven Claude skills that test a business idea before you launch, from pricing to competitors. Paste the repo address into Claude and say "install skill".
5. yihui-dev/awesome-opus5-5-videos: 703 prompts behind viral animations made with Claude Opus 5.5. Copy one prompt and run it with your own agent.
If you want the quickest result, start with the last one. A single prompt produces an animation.
Having the model grade its own conversation right after answering is cheap oversight. Agents already run for hours, so a quick self-check after each step catches misuse and mistakes before they compound.
Anthropic introduced self-assessment, a new technique that lets Claude give each conversation a preliminary safety assessment right after it responds.
> On Claude Fable 5, self-assessment flags 78% of the conversations that a stronger, prompted classifier flags as potentially harmful, and 94% of the high-priority ones, at a 1% false-positive rate.
Anthropic also shed light on how self-assessment and prompted classifiers help block attacks.
> "Each successful attack costs roughly four times as many generated tokens as with no safeguards."
We already live in a world where your security depends on the compute allocated to defend the systems you rely on.
Most of this week's AI agents cost money or require a waitlist.
Gemini Spark costs $20 a month. Claude Duet also costs $20 a month, and it's still a waitlisted preview. Grok Bot costs $30 a month. Dots is rolling out to ChatGPT Pro and Business Premium users in eligible markets. Instinct needs an invite code from a current user. Muse is free, and most people can open it tonight.
Downloadable models work differently. Nemotron 3.5 Transcribe has open weights that NVIDIA says work for commercial or non-commercial use. Mistral Large 4 is also free to download. Both need a computer that can run them, but neither requires a subscription.
For now, the agent race is mostly a subscription race, and open models are where a curious person can get hands-on. Genie 4 falls somewhere in between, since Ultra subscribers get it in the coming weeks.
If you want the cheapest test, Mistral's free demos run for the first week.
Mistral's biggest model, Large 4 ("Le Chonk"), is free to download, and its official demos are free to try for the first week.
Mistral says it is the best open-weight model for coding. On a coding agent test, it ties GLM-5.3 while using 47% fewer tokens per task.
Reading with it is cheap: about 8 cents to have it read a 300-page book. It is also fast, at about 120 words a second.
It still trails the best Chinese open models on the broad index. For coding, it's worth a try. For everything else, maybe wait.
Lower inference cost per unit of reasoning is what makes open models usable inside agents. Agents make thousands of calls a day, so a model that's cheaper to run wins over one that's slightly smarter but expensive.
MISTRAL 🔥: Le Chonk (Mistral Large 4) is now available on HuggingFace!
> "The new model also features substantial gains in reasoning efficiency that give developers greater performance while lowering inference costs."
> "Mistral Large 4 ranks as the best open-weight model for coding and among the world's most capable open models overall."
It now takes the top spot on Cognition's FrontierCode benchmark among open models!
We need new Mistral Small now 👀
Open weights for agentic coding and cyber defense is the part that matters. Teams can audit what the model does before it touches their repos or their network, and that trust is what gets agents deployed for real.
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
You can now buy a thin laptop that keeps an AI agent running in the background, starting at $2,599.
RTX Spark laptops are on sale now from ASUS, Dell, HP, Lenovo, Microsoft and MSI. They're 14mm thin and weigh three pounds, and the graphics chip and processor share up to 128GB of memory. That's enough for AI work that used to need a desktop.
In the NVIDIA App, you can pick whether your open app or a background agent gets more of the GPU. That only works with one Windows option switched on, so turn it on first.
Codex is OpenAI's coding agent, and it runs on your own computer. The short way in:
1. Install it with Homebrew: brew install --cask codex.
2. Type codex in a terminal and pick "Sign in with ChatGPT".
3. Prefer a window? Run codex app.
It works with a Plus, Pro, Business, Edu or Enterprise plan. Otherwise you need an API key.
The limit: it was built for people who write code. Start in a folder you don't mind changing.
Google Search can now act inside your apps. Connect Instacart, Canva or YouTube Music and AI Mode finishes the task on the results page.
Rolling out this week in the U.S., you can:
1. Add barbecue ingredients to your Instacart cart, then check out in the Instacart app or on its website
2. Ask Canva for flyer templates for a project
3. Build a party playlist and send it to YouTube Music, then press play
Calm is coming to both Gemini and Search. Gemini already has over 50 connected apps, including Adobe, Peloton and Experian.
Developers in the U.S. can join the beta waitlist. You connect a cloud-hosted MCP server, register and test the app in Google's console, then submit it. Once it's live, you can track traffic, latency and error rates.
If you're in the U.S., open AI Mode and try Canva for a quick test. Search is taking over the planning step, but checkout still happens in each app.
Passing secrets to agents without pasting them into the chat is the feature people actually need. Keys in prompts leak into logs and history, so secure injection beats any clever workaround.
Yes, but make it portable. Skills, MCP servers and agent instructions should live in the repo, not locked inside one tool. Whoever makes that config work across every agent harness wins.
The quote is more grounded than "RSI" sounds. AI speeding up training and launch cycles already counts, and it compounds: faster iterations mean more experiments, which mean better models, which mean faster iterations.
Open weights is the real headline here. Meeting audio is the stuff companies least want to send to a third-party API, and now it can stay on their own servers.
NVIDIA is releasing an open speech model that can transcribe a meeting, label who said what, translate it and answer questions about the recording:
Nemotron 3.5 Transcribe supports:
- Transcription in 40 languages
- Translation between English and 31 languages
- Up to eight speakers, with speaker labels in 13 languages
- Recordings up to an hour long
You can give it names and technical terms to guide the transcript, or ask questions about the audio. Audio Q&A is trained in English.
It has roughly 30B parameters, with about 4B active per token, and can run on your own NVIDIA GPU.
According to NVIDIA's launch materials, it ranks first among open-weight models on Hugging Face's English Open ASR leaderboard and leads VoiceArena's 14-system test of transcription with speaker attribution.
Built for recorded audio. Live streaming uses a separate model.
Super cool to see!
Model: huggingface.co/nvidia/NVIDIA…
Demo: huggingface.co/spaces/nvidia…
Blog: huggingface.co/blog/nvidia/n…
Alignment is the boring layer that decides if voice products feel pro or cheap. Subtitles half a second late or dubbing slightly off sync is what users notice, not the model behind it.
New from ElevenLabs research: Today we’re sharing more about Cue, our new forced alignment model. It places every character on the audio timeline with more precision than any other aligner we tested. That means the right words show up at the right time.
Forced alignment is the step that turns a transcript into timestamps. It’s how subtitles land on time, how dubbed lines match the speaker, and how clicking a word in a transcript jumps to the moment it was said. Most open-source aligners are built for short, clean clips in a limited set of languages. Cue is one model that handles recordings of any length in a single pass. On real-world audio, it was more accurate than every open-source aligner we tested in all 29 languages of our benchmark. It aligns an hour of audio in about half a second on one GPU.
Cue already powers every Speech to Text transcription on ElevenLabs, and it’s available on its own through the Forced Alignment API.
Read the full post: elevenlabs.io/blog/cue-align…
A #1 on a leaderboard is a reason to test it, not a reason to switch. Run it on your own ugly real tasks for a day and the ranking either holds up or it doesn't.
