Ad
Skip to content

Matthias Bastian

Matthias is the co-founder and publisher of THE DECODER, exploring how AI is fundamentally changing the relationship between humans and computers.
Read full article about: Sakana AI launches "Ultra Deep Research" to automate weeks of strategy work

Japanese AI startup Sakana AI has unveiled "Sakana Marlin," its first product for business customers. The system works autonomously: give it a topic, and it researches on its own for up to eight hours, then delivers detailed reports and presentations. Sakana AI says the tool can produce professional strategy analyses that would normally take human teams several weeks.

Eine Übersichtsansicht eines von der KI generierten Beispiel-Dokuments. Zu sehen sind mehrere hochformatige Textseiten eines Berichts sowie darunterliegende querformatige Präsentationsfolien in japanischer Sprache. Auf dem Deckblatt sind das Logo von Sakana AI, der Name „Sakana Marlin“ sowie das Wasserzeichen „Sample output“ zu erkennen.
Sample output from "Sakana Marlin": after autonomous research, the tool creates text reports and presentation slides on a given topic (here: AI trends in the financial sector). | Image: Sakana AI

Sakana Marlin combines the company's "AI Scientist," designed to resolve contradictions, with its previously introduced "AB-MCTS" method for strategic searches. Multiple AI models work together, and longer thinking time is meant to yield better results, the company says.

The company is looking for beta testers in finance, research, and business consulting. The beta is free, but requires registration (the form is in Japanese). The biggest weakness of automated reports like these is hard-to-spot AI errors, something the startup doesn't address in its announcement.

Read full article about: Microsoft's MAI-Transcribe-1 runs 2.5x faster than its predecessor at $0.36 per audio hour

Microsoft has introduced MAI-Transcribe-1, a speech-to-text model supporting 25 languages that achieves the lowest word error rate of any model tested on the FLEURS benchmark, beating Scribe v2, Whisper-large-V3, GPT-Transcribe, and Gemini 3.1 Flash-Lite. The model is also built to handle tough recording conditions like background noise, poor audio quality, and overlapping speech, Microsoft says.

MAI-Transcribe-1 (green) leads in word error rate on the FLEURS benchmark in most of the 25 languages tested, outperforming Scribe v2, Gemini 3.1 Flash-Lite, Whisper-large-v3, and GPT-Transcribe. | Image: Microsoft

Microsoft is rolling out MAI-Transcribe-1 across Copilot Voice and Microsoft Teams. Developers can try it as a public preview through Microsoft Foundry and the Microsoft AI Playground. The model runs 2.5 times faster than Microsoft's previous Azure Fast offering and costs $0.36 per audio hour. Combined with MAI-Voice-1 and a language model, it can also power voice agents, Microsoft says.

Cohere and Mistral recently released open-source alternatives that perform at a similar level.

Read full article about: Anthropic's leaked AI coding tool has been cloned over 8,000 times on GitHub despite mass takedowns

Following the accidental leak of its AI coding tool's source code, Anthropic has had more than "8,000 copies and adaptations of the raw Claude Code instructions" removed from GitHub via a copyright request, the Wall Street Journal reports. One programmer already used AI tools to rewrite the code in different languages, keeping it available despite takedowns. This shows just how damaging a code leak is in the age of AI: once it's out, it spreads faster than anyone can contain it.

The code contains valuable techniques Anthropic uses to control its AI models as coding agents—the "harness"—including a "dreaming" function for task consolidation. Competitors now have a blueprint to replicate Claude Code's capabilities, weakening Anthropic's edge in an already cutthroat market.

The timing is particularly bad: the company is planning an IPO at a $380 billion valuation, and this kind of leak is unlikely to sit well with investors. It also comes just days after a separate leak about Anthropic's new AI model Mythos, also caused by human error within the company's content management system.

Comment Source: WSJ

Google Deepmind study exposes six "traps" that can easily hijack autonomous AI agents in the wild

AI agents are expected to browse the web on their own, handle emails, and carry out transactions. But the very environment they operate in can be weaponized against them. Researchers at Google Deepmind have put together the first systematic catalog of how websites, documents, and APIs can be used to manipulate, deceive, and hijack autonomous agents, and they’ve identified six main categories of attack.

Read full article about: OpenAI officially confirms mega-funding round and ChatGPT super app

OpenAI has officially closed its latest funding round. The company raised $122 billion at a valuation of $852 billion. Key backers include Amazon, Nvidia, SoftBank, and Microsoft, along with a16z, BlackRock, Sequoia Capital, and several other investors. Private investors put in $3 billion through banking channels, and the company also expanded its credit line to $4.7 billion.

OpenAI says it's now pulling in $2 billion in monthly revenue and has crossed 900 million weekly active ChatGPT users. The company also officially unveiled the ChatGPT Super App, a single product that rolls together ChatGPT, the Codex coding agent, web search, and what OpenAI describes as "our broader agentic capabilities into one agent-first experience."

The bulk of the new capital will go toward computing infrastructure. OpenAI is clearly leaning harder into enterprise going forward; the company recently killed off its Sora video model to free up compute and because it wasn't gaining traction anyway. Enterprise already accounts for more than 40 percent of the company's total revenue. "Our consumer scale becomes the front door for enterprise usage, as familiarity in daily life drives adoption at work," OpenAI writes.

Read full article about: Oracle reportedly lays off thousands of employees to bankroll its massive AI infrastructure bet

Oracle is laying off thousands of employees to fund its AI infrastructure push. Business Insider broke the story, and CNBC confirmed the cuts through two anonymous sources. Oracle, which had 162,000 employees as of May 2025, declined to comment.

The cuts stem from Oracle's aggressive AI spending, which has pushed the company into debt as cash flow shrinks. Since announcing plans to raise $50 billion in January, the stock has lost roughly a quarter of its value. TD Cowen analysts estimate that eliminating 20,000 to 30,000 positions could free up as much as $10 billion in cash flow.

On an earnings call, co-CEO Clay Magouyrk defended the spending, saying AI hardware demand outpaces supply. He pointed to $553 billion in guaranteed revenue, including a $455 billion order from OpenAI. But whether OpenAI can actually pay up remains unclear; the ChatGPT maker is also burning through cash at a rapid clip.

Oracle's internal termination email cites only "current business needs" without giving a specific reason. Meta is also reportedly planning large-scale layoffs to offset its own massive AI infrastructure costs.

Read full article about: Anthropic accidentally publishes Claude Code source code for anyone to find

Anthropic inadvertently published parts of the source code for its AI coding tool, Claude Code. Developers discovered more than 500,000 lines of source code and over 1,000 related files on NPM, a public repository where developers share JavaScript software packages. When publishing Claude Code as an NPM package, Anthropic accidentally included far more internal files than intended, including details about how the tool works and references to unreleased models and features.

Anthropic says the leak was caused "by human error," not a security vulnerability, and that no customer data was affected. The company is working on measures to prevent similar incidents. This is Anthropic's second leak in just days, coming right on the heels of internal blog posts about their new Mythos AI model accidentally slipping out.

Read full article about: Google's Veo 3.1 Lite cuts video generation costs by more than half

Google Deepmind is launching Veo 3.1 Lite, its most affordable video generation model yet. It costs less than half the price of Veo 3.1 Fast but matches its speed, according to Google. The company doesn't specify quality differences between the three tiers.

Veo 3.1 Lite supports text-to-video and image-to-video at 720p and 1080p in portrait and landscape formats, with clips of 4, 6, or 8 seconds. Pricing starts at $0.05 per second for 720p. Starting April 7, Google is also dropping Veo 3.1 Fast prices.

Pricing (per second in USD) Veo 3.1 Lite Veo 3.1 Fast Veo 3.1
720p $0.05 $0.15 ($0.10 from 4/7) $0.40
1080p $0.08 $0.15 ($0.12 from 4/7) $0.40
4K - $0.35 ($0.30 from 4/7) $0.60

Veo 3.1 Lite is available now through the Gemini API and Google AI Studio, with more video news for developers coming soon. Full details are in the developer documentation.

Google's announcement comes right after OpenAI revealed it's shutting down Sora. That leaves Google facing serious video generation competition primarily from China, especially Alibaba's Seedance 2.0, which delivers higher quality but comes with copyright concerns.