AI from Malta 🐝 Voice agents, chat agents and workflows for European businesses. 15–30 hours back in month one. Foundation in 14 days. https://xitter.cf/t.co/mhCA6yEaPf
Malta
Joined March 2026
- Tweets400
- Following77
- Followers19
- Likes784
Pinned Tweet
Every business runs on conversations. Most go unanswered.
Yours won't.
Maya answers the chats. MAX answers the calls. That's his real voice in the film.
beesofai.com
Bees Of AI retweeted
Voice apps have pretty strict requirements for latency, privacy, and customization (understandably).
So it's often hard to find an AI voice tool that really fits your needs.
In this tutorial, @juntao helps you build your own voice AI agent using open source tools.
freecodecamp.org/news/how-to…
Two desktop machines holding a terabyte of unified memory, running an agent class model on site. For clinics, firms and front desks with sensitive records, where the model runs may soon matter as much as how well it scores.
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
Meta released open weights for Muse Spark Open: 816B parameters, 42B active, a 1M token context, built for coding and agent work. It runs on two Mac Studios. A frontier class agent that never leaves the building changes the privacy conversation for real teams.
SpaceXSI claims under one second of latency, speech to speech. Independent benches now score it on support-call realism. For voice agents the useful bar is resolved calls under real interruptions, not clever demos.
SpaceXSI shipped Grok Voice 3.0, a speech-to-speech model. Artificial Analysis ranks it first on its Speech-to-Speech Index at 86.9, with 97% on Full Duplex Bench. Voice agents are now being scored on how they handle support calls, not on how they sound in a demo.
1/ Anthropic published new numbers on distillation: covert campaigns that pull a frontier model's capabilities into another model without permission, often through networks of fake accounts.
6/ After the classifier, Anthropic says reasoning tokens distillers could extract on its platform fell by a factor of 10 to 100. Harvested data down around 90 to 99 percent. Tens of thousands of accounts banned, often within two hours of a campaign start.
What distillers want most is the reasoning transcript, not the reply. If that holds, guarding how a model thinks is the real control plane. Anthropic says harvest on its platform fell by a factor of 10 to 100.
Over the last few months, we’ve made significant progress in reducing the effectiveness of distillation attacks against Claude.
In this post, we explain how: anthropic.com/research/disti…
Anthropic says new defenses cut the amount of Claude reasoning data distillers harvest by around 90 to 99 percent. The prize is not the final answer. It is the thinking transcript that teaches a smaller model how Claude got there.
Tokens per megawatt is the metric that matters once chat and agents run all day. Silicon only earns its keep if that number holds in production.
Today we are announcing our biggest bet yet on our own silicon.
Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services.
And we’re introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems.
We are already deploying Maia 200 across the US, Europe, and Australia, delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot.
Our goal is to build on the full-stack efficiency gains we’ve already seen, while continuing to bring the latest innovations from across the ecosystem into our fleet.
Microsoft announced Azure Maia 300 yesterday: third-gen inference silicon, production in 2027, primary for its own AI products.
Built for long-context reasoning, latency-sensitive chat, and multi-step agents. Claim to watch: 40% more tokens per megawatt than leading GPU systems.
AMIE's list of possible diagnoses included the final one in 90% of those 98 visits. Interesting, and still not a green light for unsupervised care.
The study was feasibility under live supervision. Google's own note says larger trials are still needed.
A chat that gathers history before the visit is only useful if a person still owns the room, and if a live supervisor can stop it.
Our medical research system, AMIE, is the first patient-facing conversational diagnostic tool of its kind to be studied prospectively in a real-world clinical setting.
In a study, published today in @TheLancet, we found that when patients chatted with AMIE before an in-person appointment, the interaction built their confidence and helped them organize their thoughts. And physicians who reviewed the information from the patient conversation got more time back to focus on collaborative care and shared decision-making instead of digging through data.
Google and Beth Israel Deaconess put AMIE, a research diagnostic chat agent, in front of 98 real patients before urgent care visits. A physician watched every chat live. None needed a safety stop.
The Lancet study says the summary helped clinicians prepare in 75% of cases.