Ein kuratierter TLDR-Stream fuer deine Themen.

Without large amounts of compute, small labs can only compete through algorithmic efficiency. Magic's pretraining recipe is now more than 10 times more compute-efficient than that of leading open-weight base models. The startup believes that pretraining, agentic RL, and long-context are sufficient for building superhuman coding agents and automating AI research and development. This post discusses its pretraining and long-context work.

OpenAI plans to introduce Managed Agents at DevDay 2026, following a model similar to Anthropic's offerings, targeting businesses and developers. The new agents aim to integrate advanced models with superior computer-use capabilities at competitive price points to attract users. Additionally, they will offer functionality that enhances advertising through interactive agents, potentially challenging major players like Meta and Google.

AI makes software cheap enough to customize around individual workflows, interests, and aesthetics, shifting value from standardized apps toward malleable personal tools. Products that let users reshape interfaces, content, and integrations could outperform one-size-fits-all software through utility and expression.

Anthropic's Economics team developed a model predicting AI's impact on the US economy through various scenarios ranging from moderate growth to extreme transformations. In scenarios of rapid AI-driven growth, knowledge workers face greater unemployment and wage stagnation, while non-knowledge sectors may see increased wages. The model highlights economic inequalities, with more gains potentially flowing to capital rather than labor.

Automation has been good, but AGI may make humans economically useless, which will have many consequences. It's clear that AI capabilities are growing far faster than our ability to control or understand them. In the future, people will likely hand over control to AI voluntarily, as it will be the logical choice. However, before that happens, there are several signs that people are already giving up thinking.

Automated AI text detection is currently an underserved niche. Pangram does an excellent job, but it is still more a tool to confirm suspicion. This developer wanted a tool that runs in the background and automatically scans sites so they can avoid AI-generated text in the first place. This post details how they developed Deckard, a Chrome extension that uses a locally-run model to mark AI-generated text.

Frontier labs may be applying probabilistic AI safety techniques to problems that require deterministic security controls. Recent agent sandbox escapes highlighted the gap between reducing harmful model behavior and reliably containing software.

Meta introduces Muse, a personal AI agent, powered by Muse Spark, to help users achieve goals by automating tasks like booking travel or sending emails. Muse operates securely on Muse Secure VM, ensuring data privacy with unique protections like the Sentinel agent overseeing actions. Muse will soon offer encrypted data with Muse Confidential VM and is available on iOS, Android, and muse.ai in the US.

Andrew Tulloch, one of the highest-paid employees in the tech industry, had worked in Meta's TBD lab.

Claude successfully created the first complete computer-verified proof of Fermat's Last Theorem in 11 days using Lean, automating the complex task initially proven manually by Andrew Wiles in 1995. The proof, verified via Prove2Me and Lean, involved 13 million lines of code and proved 29,500 intermediate theorems, proving AI's potential to ease the traditionally laborious formal verification of mathematical proofs.

Build an ICP, give AI persistent business context that doesn't need to be re-explained every session, and audit content for risk reduction rather than just engagement.

A growth engineering intern at Browserbase built an AI engine to automate its SEO and AEO workflow. 53% of search traffic came from just five posts, while only 17 of 77 posts had been cited by AI engines. BlogEO audits existing content, finds search opportunities, generates drafts, and tracks AI citations. Human approval remains in the loop. After 6 weeks, search impressions grew 5.8X, page-one queries grew 9.8X, and average blog position improved from 10.9 to 6.6.

AI creates little value when it simply speeds up a broken process. Companies should redesign the workflow first, then use software, agents, and humans where each is best suited.

Tencent's TeamAI CLI distributes a team's skills, rules, MCP configuration, and project knowledge across tools such as Codex, Claude Code, Cursor, and OpenCode. Its shared Git-based harness keeps agent guidance synchronized and adds optional retrieval, learning capture, and codebase knowledge features.

AI labs are releasing model updates at a faster pace, making it harder for businesses to evaluate which tools are worth adopting. Anthropic, Meta, Google, and OpenAI all launched updates in one week. The rapid rollout is creating “model fatigue” as teams spend more time comparing capabilities and performance.

AI models are now capable of taking a request from Slack and then building a feature from end-to-end, with humans just having to click 'merge' or 'deploy' at the end. There's not really much left for humans to do. Humans are still better at absorbing, parsing, and sharing context with each other. However, once that can be automated, the loop will finally be closed, and there will be nothing left to do but to pay the bill.

The price of AI is coming down. However, volume growth may not be large enough to counteract the effective decline in prices. The models driving volume increases are relatively cheap, standard models. These trends do not appear to be driven by the adoption of open-source models or Chinese models.

Modern cancer-vaccine design involves identifying mutations, predicting presentation, predicting immunogenicity, ranking targets, and designing RNA. However, biology is complicated, and whether a target makes a good vaccine candidate depends on all kinds of information. This makes cancer vaccines an unusually direct test of whether AI models can take what it has learned from biology and apply that knowledge to design a medicine for one particular human being.

PI-Desktop is an early-preview desktop workspace for coding agents that keeps projects, sessions, reviews, and permissions in one local-first app. It supports multiple model providers, MCP, skills, subagents, and plugins, while routing privileged work through a reviewable permission layer.

Next-token prediction optimizes for the shortest path to an answer, not necessarily the best or safest one. Prompts and safeguards cannot fully solve this architectural limitation, and it's worth exploring alternatives such as world models.

This roundup argues that frontier-lab coding agents are already performing substantial research and engineering work alongside their creators. It pairs claims about rapid growth in multi-agent usage and benchmark performance with a reminder that public concern about AI remains high.

Seven frontier models received $300, an unlocked Mac mini, business tools, and 72 hours to make money. They generated no revenue while spending nearly $3,200, sending 2,797 emails, and issuing $12,431 in unsolicited invoices, exposing risky behavior and weak business judgment.

Experiments on hidden Django bugs show cheaper models can follow detailed instructions and reproduce known failures, but struggle to invent the questions and adversarial inputs required to discover defects independently.

Omarchy is an opinionated Linux distribution built for AI agents.

AI makes execution cheaper, so a design system's value increasingly comes from preserving decisions instead of merely supplying components. Clear semantic names, fewer overlapping choices, and encoded institutional knowledge prevent people and agents from multiplying ambiguity across a product.

This article surveys recent work on recurrent-depth transformers, where the same blocks are reused across multiple passes to reduce parameter memory without reducing computation. The piece also examines hidden reasoning and what these architectural choices imply for the next generation of language models.

Meta describes an AI “second brain” built from structured, auditable knowledge files rather than retrained model weights. The system separates facts from procedures, turns expert feedback into regression-tested updates, and reduces assessment work from days to minutes while keeping changes transparent and reversible.

A developer describes the strange shift from flow-state coding to orchestrating 20-plus agents, and why the software factory itself is becoming the product.

When the same prompt is sent to an AI model multiple times at zero temperature, variations in the results can occur due to the nondeterministic nature of floating-point arithmetic combined with server load dynamics. The model's output can differ based on how requests are batched and how the GPU processes the addition of numbers, leading to discrepancies despite identical settings.

OpenAI appears to be implementing a new feature that helps ChatGPT write in users' voices by referencing examples from their connected apps. Some users have had the feature for weeks, but the company has yet to officially announce the tool. OpenAI often tests new ideas with select users ahead of launch. The article includes a screenshot of the feature interface.

Spotify routed large file reads and predictable code generation from Claude Code to cheaper Gemini 2.5 Flash workers through Portal modes and a Claude Code plugin. Hooks enforce delegation for files above a size threshold, producing about 90% mean token savings in four Java monorepo scenarios while keeping debugging, architecture, and editing with the frontier model.

Hunter Biden is launching $LAPTOP on Coinbase's Base network on September 9. The token name references the 2020 laptop controversy. The tokenomics earmark 20% of the supply for investors who lost money on $TRUMP and include a conditional burn mechanism tied to 30 predetermined outcomes, including a Democrat winning the 2028 presidential election, Bitcoin reaching a new all-time high, and $LAPTOP's market cap surpassing Solana's. Up to 30% of total supply faces burns if all conditions are met, while missed targets redirect proportional token amounts to charity. The structure embeds outcome-based supply changes that place it closer to a prediction market instrument than a conventional memecoin.

AI makes execution cheaper and experimentation faster, but product expertise still matters. The advantage shifts toward judgment, customer understanding, taste, and verification.

Product managers should build engineering understanding with AI as product and engineering roles converge.

AI agents have created "slop-creep" - the accumulation of unnecessary features, abstractions, and infrastructure built simply because they are cheap to produce. Without implementation friction, weak ideas survive longer, especially in large companies where software persists.

AI vendors are increasingly marketing behavioral science, emotional calibration, and conversation design as core capabilities.

Evals tell you whether AI works, not whether it creates value.

AI can increasingly handle execution, but humans still choose the next hill to climb. Consumer AI's biggest opportunity may be helping people feel happier and more connected, not merely more productive.

Tailwind is joining Shopify, one of the first companies operating at scale to see the potential of Tailwind CSS and start building with it. Shopify has invested in making sure Tailwind is actively maintained and will continue to provide a stable long-term home for the project. Nothing will change with Tailwind CSS or any of Tailwind's other open source projects, but the project will no longer try to grow its business. All existing customers will maintain access to their products, but Tailwind has closed signups for new customers.

Microsoft's CoreAI portals all sat on Fluent 2 and still felt disjointed, because shared components did not prevent teams from inventing different navigation, layouts, and flows. The CoreAI Design System, forked from Foundry's work, ships guidance through both a Figma library and an experimental MCP server so coding agents can consume it too.

Project Oasis is Adobe's new web-based graphic design app, described as a modern workspace with brand-aware AI built into the workflow. A September 7 community post invites agency, freelance, and in-house designers into early access, with a private feedback channel and an NDA. Pricing, a full feature list, and a launch date remain undisclosed.

Apple has unveiled the iPhone Duo, its first foldable. Featuring a 7.6-inch inner display, a 5.4-inch outer screen, Touch ID, and an adaptable iOS 27 interface designed for split-screen and half-folded use, the wider device supports paired apps and tabletop video. Counterpoint estimates it could capture up to 25% of the foldable market by year-end, despite foldables accounting for less than 2% of global smartphone shipments.

AI is becoming a useful part of UX work, but it works best as an assistant rather than a replacement for human judgment, research, and design expertise. The strongest use cases are targeted ones—speeding up research planning and synthesis, prototyping, writing, ideation, workshops, and service design—while keeping real users, critical evaluation, and clear accountability at the center. As UX roles evolve, the most durable skills are shifting toward strategy, critique, taste, systems thinking, and the ability to curate meaningful context for AI rather than simply producing artifacts faster.

Six principles set the quality bar for a design engineering practice being built at DuckDuckGo, where the team is still small and hiring its second member. They cover understanding problems before making, owning the experience past handoff, making trust effortless, reducing scope rather than quality, finishing every state, and making craft repeatable. Moving fast by making less, better is the one expected to be tested most often.

Vercel's design.md is a public file any agent can load to build on-brand pages outside the codebases where its product-design skill lives. The system pairs that guidance with a public stylesheet defining fixed classes and tokens, plus an evaluation loop that turns reviewer feedback into rules and deterministic checks. After more than 200 test runs, pages generated with design.md loaded showed 57% fewer known layout failures than those generated without it.

Solaris is the first Interface World Model, generating apps and websites frame by frame as a user interacts, with no code layer in between. A study of 250 participants preferred it over a coded interface in 61% of instruction-following comparisons and 71% of those judging natural behavior. Legible text, trust in generated content, and coherence across long sessions remain unsolved.

Lora Zombie, a self-taught painter, built an international following through DeviantArt before exhibiting in Toronto, Los Angeles, London, and New York.

ByteDance is developing a real-time spatial video model, built on its Seedance system for Pico headsets. The model would generate interactive 3D worlds in the cloud at around 20 frames per second with roughly 50 milliseconds of latency, shifting rendering costs away from device hardware. A launch could come as soon as next month, part of ByteDance's broader 2026 push to lead in world models alongside $30 billion in secured loans and up to $70 billion in planned AI infrastructure spending.

Code discovered in the latest iOS 27 beta suggests Apple may bring advanced photography features to the Camera app, including exposure histograms, waveform monitors, focus peaking, zebra highlight warnings, and manual focus controls. The update could also add burst-style self-timer shots and image processing tailored for the rumored variable-aperture camera on the iPhone 18 Pro. If released, these features would significantly narrow the gap between Apple's built-in Camera app and professional third-party camera apps.

HBO's simplified Hogwarts house crests for its upcoming Harry Potter series have divided fans.

NN/g research finds real differences between carets, arrows, and plus icons on accordions, with the caret the safest choice and the most tapped. With no icon at all, users expect the item to open a new page. Split menus, where label and icon trigger different actions, cause the most mistakes and are best avoided.

Archyl maps every system, service, and dependency into an interactive C4 model, then keeps it in sync on every push.

Google's Gemini Spark can now edit images, curate albums, auto-create shared albums, and turn concert flyer photos into calendar entries inside Google Photos. Google says the rollout arrives over the coming weeks to eligible Gemini AI Pro and Ultra subscribers in the US, though international timing remains unannounced.

World Labs has introduced Atlas, a next-generation world model for spatial intelligence, pretrained from scratch as a multimodal autoregressive diffusion transformer spanning text, images, video, and 3D. Atlas generates up to one minute of camera-controlled 1440p video, reconstructs real-world scenes from just two or three images, and simulates space and time for robotics. It outperforms specialized 3D reconstruction models on benchmarks and will power future versions of Marble and other World Labs products, now entering early access with select partners.

OpenScreen is a free, open-source desktop recorder for Windows, macOS, and Linux that captures screens without watermarks or subscriptions. It uses GPU-accelerated recording and is designed for quickly producing product demos and shareable videos.

JetBrains is urging users of Cadence, its hosted cloud computing service, to revoke and rotate all credentials after unidentified attackers exploited CVE-2026-63077 (CVSS 9.8) to breach its own environment. The vulnerability stems from deserializing untrusted data and allows an unauthenticated attacker with access to a TeamCity server to bypass authentication checks and execute arbitrary operating system commands with the privileges of the TeamCity server process. Attackers reached a 2024 Cadence backup holding AWS IAM credentials, S3 files, PyCharm-synced source code, and personal data, so defenders should treat all executions as untrusted and audit connected AWS accounts, registries, and repositories for activity since August 8.

German software company SAP has released an update that patches 20 security vulnerabilities. Of particular note is a maximum-severity buffer overflow vulnerability in the External Passport Protocol (EPP) processing library, which could allow an unprivileged attacker to run arbitrary commands as an administrator. Another missing authentication vulnerability in the SAP NetWeaver Message Server could allow unauthenticated attackers to execute malicious payloads and arbitrary commands across the full SAP system cluster.

There is a malware campaign targeting F5 BIG-IP APM devices that exploits a critical RCE vulnerability and injects a fileless web shell. A distinct installer deploys the payload by infecting the Apache executable and modifying SELinux configurations. ESET researchers identified the malware as PoisonedRefresh.

Threat actors are actively exploiting CVE-2026-75650, a max-severity unauthenticated code execution vulnerability dubbed StyleSmuggler, to compromise online stores running Adobe Commerce and Magento Open Source. Exploitation began on September 4, with the deployment of a Rust backdoor and a separate PHP web shell. Adobe released emergency hotfix VULN-393411 to address the flaw, prompting Sansec to warn that patching alone is insufficient for stores that may already be compromised.

Threat actors are chaining CVE-2026-81578 and CVE-2026-82078, an authentication bypass and remote code execution pair in PaperCut, to steal credentials from K-12 schools and major universities across the US and Europe. Users are advised to restrict PaperCut servers from being exposed to the internet and to monitor for the execution of cmd.exe, powershell.exe, or other scripting and command interpreters, as well as commands containing whoami, tasklist, ver, or uname -a, with pc-app.exe as the parent process.

A collection of docs, resources, and samples for implementing secure, privileged IAM in Microsoft Azure and Entra.

Security researchers gained access to an Elasticsearch cluster containing an Advanced Passenger Information System (APIS) database with records belonging to 220M passengers and flight crew members. The researchers received a 401 when trying to access the cluster via the public internet but could reach the database, which accepted default credentials via a cloud-based access path. The data includes passengers' and crew members' names, dates of birth, sex, nationalities, passport or travel document numbers, document expiration dates, and issuing countries.

Sophos is developing Exploit Path Verification to assess whether attackers can actually exploit vulnerabilities given an organization's network access, privileges, patches, and security controls. The capability will use OpenAI cyber models with analyst review, but availability has not been announced.

The G7 Cyber Security Working Group and CISA issued a joint advisory urging organizations to begin migrating to post-quantum cryptography now.

Google's GTIG reports that threat actors have graduated from basic prompting to agentic AI workflows that autonomously manage scanning pipelines, resolve errors, and execute multi-stage attacks with minimal human oversight, compressing response windows from days to hours. For instance, groups like UNC6780 exploit the AI supply chain by stealing GitHub Actions tokens to sign malicious packages by tricking LLMs, state-sponsored groups use AI across the full attack lifecycle from reconnaissance to post-exploitation, and financially motivated actors built credential-harvesting agents that stole thousands of secrets in under six hours using only AI coding tools and markdown playbooks.

Calif has demonstrated a zero-click worm that takes over a WeChat account via an unanswered VoIP call, then uses that account to call the victim's contacts. This exploit relies on a memory corruption bug in WeChat's VoIP stack and requires the attacker to be on the victim's friend list, which is a weak barrier since one compromised friend can lead to further exploitation. Tencent has mitigated this issue in server-side updates and in client versions Android 8.0.77 and iOS 8.0.76, so users should enforce these updates and reconsider trusted-contact privileges in messaging apps.

CVE-2026-6471 affects PostgreSQL 9.4 through 18. A low-privilege REPLICATION account could supply an unchecked output-plugin path, load attacker code, gain PostgreSQL superuser access, and establish persistence, while Windows targets could fetch a DLL over SMB. It is recommended to patch PostgreSQL, remove unused replication privileges, restrict pg_hba.conf, block outbound SMB and NFS, and monitor suspicious replication slots.

Hetty is an HTTP toolkit for security research. It aims to be an open-source alternative to commercial software like Burp Suite Pro, with powerful features tailored to the needs of the infosec and bug-bounty community.

Ohio man James Strahler received the first Take It Down Act conviction, with a 15-year sentence for distributing 3,000+ real and AI-generated child sexual abuse images and cyberstalking victims.

White-hat hackers drained ~$320 million from Liquid Network's federation wallet over the weekend.

Cursor now lets cloud agents run on your own infrastructure, while Cursor still handles the agent loop, inference, and orchestration. Self-hosted workers can scale dynamically across AWS, Vercel, Modal, Cloudflare, Macs, etc., which gives teams direct access to internal systems, custom hardware, and existing build environments.

Claude for Microsoft 365 integrates Claude into Excel, PowerPoint, Word, and Outlook with shared context across apps. Claude for Excel, PowerPoint, and Word is now generally available on all paid Claude plans, while Claude for Outlook is in beta. The offering emphasizes reviewable edits, use of customer templates, and deployment through existing cloud and compliance frameworks.