X / Twitter
Swyx, AI Tool Builder
Swyx tested GPT Luna Max and Claude Fable Ultracode with the same prompt, asking both to recreate Grok Imagine using fal's open models. Fable achieved better visual fidelity, but Luna understood the intention to use open models more accurately and produced a more useful result. This shows that interface similarity and understanding task intent are separate evaluation dimensions; screenshots alone cannot establish the quality of an agent's work. He also mentioned that PDB environments now experimentally support AFS-like cloning without restrictions on runtime or programming language. He envisages turning every Git command into an agent-native operation, adapting version control directly to automated development workflows.
https://x.com/swyx/status/2087045848022843451
https://x.com/swyx/status/2087017780617126075
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
Thibault Sottiaux announced that usage allowances had been reset for all paid ChatGPT Work and Codex users. A more substantial update came in cybersecurity: OpenAI is expanding access to frontier cyber capabilities through new Daybreak Blue and Red access tiers. The company also introduced GPT-5.6-Cyber for defensive tasks such as vulnerability discovery, patching, and penetration testing. For organizations unsure where to begin, he recommended contacting partners with access to the latest cyber models for help investigating and rapidly fixing issues. These updates show frontier models' security capabilities moving from general access toward deployment that combines tiered authorization with professional services.
https://x.com/thsottiaux/status/2086972933566857393
https://x.com/thsottiaux/status/2086874565909815403
Peter Yang, AI Tutorial Author
Peter Yang summarized five lessons from Linear's experience building production-grade agents. First, start by reconstructing the real workflow rather than writing a prompt: establish where work begins, which systems hold context, what completion means, and where human review is required. If work starts in Slack, use Slack as the entry point instead of forcing users into a separate chatbot. Second, do not put all context into the prompt; give the agent retrieval tools so it can load what the current task requires. Third, solve one frequent job first, then expand based on real usage. Linear initially only converted sales notes and Slack discussions into issues. Fourth, use the strongest model while the process is still being established; after defining success criteria and evals, test whether smaller models can handle narrow tasks. Fifth, turn every failure into an eval or product task: if the tool exists but behavior is wrong, add an evaluation; if a tool is missing, record a gap in system capabilities.
https://x.com/petergyang/status/2086824976800436676
Madhu Guru, Senior Director at Meta AI
Madhu Guru is exploring a central question at the intersection of AI and consumer experiences: how can products understand why a user takes an action, rather than just recording what they did? Search and chat provide explicit expressions, while watching, skipping, lingering, and returning are implicit signals. Interpreting these behaviors requires considering users' life circumstances, events in the outside world, and interests that change over time. The real difficulty is performing this reasoning nearly in real time in products serving billions of users. His focus goes beyond traditional recommendation systems' behavioral statistics toward dynamic modeling of intent, context, and evolving interests.
https://x.com/realmadhuguru/status/2086909974668784113
Thariq, Anthropic Claude Code Team
Thariq believes collaborating with AI requires two key abilities. The first is allocating compute, because most work does not come with a prewritten list of the "most important problems"; people must decide which problems merit computational resources. The second is being a thinking partner: users need to truly understand the problem and results to verify whether an AI's proof or answer holds up. Both abilities depend on deep technical experience and intuition, rather than simply handing a task to a model. He hopes the eventual outcome is that people retain technical depth while progressing faster on important problems.
https://x.com/trq212/status/2086931647468097932
https://x.com/trq212/status/2086931648898342914
Google Labs, Google's AI Experiments Team
Google Labs announced that the Portraits experiment will end on September 14. By launching experiments quickly, collecting user feedback, and continuously learning, the team gained extensive insights into expert-grounded AI. Although Portraits itself is ending, that knowledge will not be discarded. Google Labs said it will incorporate the results into other Google products. This reflects the Labs experimentation model: an individual product can have a clear endpoint, while validated interactions and technical lessons move into a larger product ecosystem.
https://x.com/GoogleLabs/status/2086936798710923603
Guillermo Rauch, Vercel CEO
Guillermo Rauch said DeepSec has become a routine part of Vercel's internal cyber defense work, with the team even using "deepsec" as a verb and command. He described it as a deep code-security review tool and believes it should be an essential stage in software factories. He also emphasized that agent sandboxes must isolate both compute environments and networks. Kimi's paper showed that traditional container isolation can suffer kernel panics and deadlocks when facing frontier models, so Vercel Sandbox uses microVMs to strengthen compute isolation. The network side is equally important: in a case disclosed by OpenAI, a model obtained internet access through an unknown Artifactory vulnerability. Vercel has made its egress firewall free, allowing developers to further restrict agents behaving abnormally. His central view is that security boundaries cannot rely on containers alone; runtime, credentials, and outbound network paths must all be controlled.
https://x.com/rauchg/status/2086965425968148806
https://x.com/rauchg/status/2086946535716393209
Aaron Levie, Box CEO
Aaron Levie sees US companies releasing frontier-capable open-weight models as an important turning point in AI. In his view, the open-weight release of Muse Spark 1.2 provides a practical way for the United States to compete in open models. Enterprises can deploy models locally or in private clouds, opening use cases in heavily regulated fields such as law and healthcare that previously struggled to adopt AI. Developers can also post-train for vertical tasks to achieve domain performance general models do not directly provide. Open weights strengthen sovereignty and continuity of model supply, reducing risks to critical operations when a model is withdrawn from the market. For applications, more model families allow task-based routing and flexible combinations of capability, cost, and control. Closed frontier models will still see widespread adoption for their ease of use and strong capabilities, but open weights will lower the cost of intelligence and create more room to build at the harness layer.
https://x.com/levie/status/2087009941806797206
https://x.com/levie/status/2086802472950239618
Ryo Lu, Product Designer
Ryo Lu announced his departure from Cursor and reflected on a decade in San Francisco's technology scene. He described Cursor as the clearest embodiment of that environment: fast, intense, and ambitious, surrounded by people trying to bring the future closer. Leaving is not a rejection of that experience; he wants a slower pace, different weather, and more culture and authentic human connection in everyday life. He is moving to Asia for a fresh start, hoping to stay grounded, experience life more fully, and build products more freely. He plans to keep sharing the changes and discoveries he encounters along the way.
https://x.com/ryolu_/status/2086854498639822942
Matt Turck, FirstMark Capital VC
Matt Turck used four technology cycles to describe a problem that has never disappeared. In the Big Data era, people said the data science models were good but the underlying data was the problem. In the modern data stack era, the dashboards were good but the underlying data was still the problem. In the Gen AI and agentic AI eras, chatbots and agents appear more powerful, yet failures still often originate in data foundations. This comparison reminds teams that advances in models and interaction layers do not automatically resolve data quality, structure, or accessibility. However the product packaging changes, underlying data remains a shared constraint on AI systems' reliability.
https://x.com/mattturck/status/2086882606638153882
Zara Zhang, Independent Builder
Zara Zhang shared observations from Beijing's AGI Bar, where customers can use unlimited DeepSeek tokens for free while vibe coding and drinking beer called "AGI bubble." The bar also sells an annual Drinking Plan and displays AI company job openings on screens, combining development, consumption, and job information in one space. She also offered a method for learning design with Codex: give it a well-designed website, ask it to analyze why the design works, then request a full-page screenshot with explanations added directly to the image. Visual annotations place the analysis alongside the interface, reducing back-and-forth between reading explanations and viewing the work. She believes breaking down real examples is more effective than learning abstract theory alone.
https://x.com/zarazhangrui/status/2086838277701882031
https://x.com/zarazhangrui/status/2086758509979316423
Claude, Anthropic's AI Assistant
Anthropic announced that Claude Sonnet 5's launch pricing will become permanent. When the model launched in June, input cost $2 per million tokens and output cost $10 per million tokens. This pricing was originally scheduled to last only until August 31, but the planned increase at expiration has now been canceled. For teams already building around Sonnet 5, this reduces uncertainty about near-term changes to their cost models. Developers can continue estimating long-term usage costs at the current input and output rates.
https://x.com/claudeai/status/2086891169217122586
Official Blog
How we contain Claude across products
Anthropic believes that as agents gain permissions capable of affecting internal services, safety goals must go beyond reducing the probability of model mistakes to limiting the maximum damage from a single failure. Claude Mythos Preview was not released in April 2026 because of its potentially excessive blast radius, illustrating that sufficient capability does not mean deployment conditions are ready. Human approval of each action is also unreliable as the sole defense: Claude Code users approved approximately 93% of permission requests, and frequent prompts gradually erode vigilance. To reduce this approval fatigue, Claude Code introduced auto mode to handle safer approvals automatically, but any probabilistic defense will still miss some cases.
Anthropic divides agent risks into user misuse, anomalous model behavior, and external attacks, and organizes defenses across the runtime environment, model layer, and external content. Runtime environments impose hard limits through process sandboxes, VMs, file-system boundaries, and egress controls—for example, keeping credentials out of the sandbox entirely. The model layer uses system prompts, classifiers, probes, and training to influence behavior, but cannot guarantee complete effectiveness. On the Gray Swan Agent Red Teaming benchmark, Claude Opus 4.7 had a single-attempt prompt-injection success rate of approximately 0.1%, rising to approximately 5% to 6% after 100 adaptive attempts; Claude Code auto mode can intercept approximately 83% of over-eager actions before execution. Content from MCP servers, third-party plugins, and web search also requires separate protection, because even a reviewed connector may deliver a compromised README into the model's context. The core principle is containment: limiting what an agent can actually access and change to engineer away the worst outcomes, rather than assuming models or users never make mistakes.
https://www.anthropic.com/engineering/how-we-contain-claude
Podcasts
No Priors — Building an Autonomous Enterprise for Real-World Services with Netic Founder Melisa Tokmak
Key takeaway: AI's greatest value in real-world services is not helping companies employ fewer people, but connecting customer demand, operating rules, and frontline labor in real time to create new revenue they previously could not capture.
Melisa Tokmak is Netic's founder and CEO, formerly an engineering leader at Scale AI who also worked on go-to-market and has experience at Meta. Netic serves essential services including HVAC, plumbing, electrical work, automotive, pet services, hospitality, and consumer wellness. Its customers include large enterprises with billion-dollar annual revenues that directly reach millions of end users.
First, an agent for real-world services cannot merely be a voice bot answering calls. If a customer's heating fails at 20 degrees below zero, the system must not only respond to a call, text, or website request but also identify the home and equipment type, read historical records, decide whether to visit today or tomorrow, and match a technician skilled in boilers or newer systems. Scheduling may also consider customer lifetime value to determine who should receive the most suitable service first when labor is limited.
Second, these businesses are often constrained by mismatches between fluctuating demand and customer support availability. Operations may start at 4 or 5 a.m., with technicians beginning at 6, before support staff arrive. A resignation or absence that day can quickly cause requests to pile up. Industries such as HVAC are highly seasonal, with heat waves, cold snaps, and holidays determining key annual revenue windows. If a business does not respond immediately, customers move to the next result on Google, an LLM, or an aggregator. Netic often enters as an overflow solution, but more than 70% of its customers now use an AI-first approach, having Netic agents handle every customer's initial interaction.
Third, Tokmak rejects defining AI's commercial value solely as cost reduction. Many private equity teams begin by discussing spending cuts, but she is more interested in generating net new revenue and reinvesting in the blue-collar labor that actually delivers services. She said directly: "If we're using AI just to cut costs, that would be very sad." To change executives' thinking, products must demonstrate concrete examples of additional business, rather than remain at the level of efficiency promises.
She also sees AI lowering barriers to education and knowledge resources, giving people who previously lacked access to mentors feedback whenever they need it. But access to resources does not mean people will take action: technology can make choices easier, not make them on an individual's behalf. A more promising framing is ensuring people get timely service on the day they most need help, shifting AI's public narrative from disappearing jobs toward positive impacts on education, health, and real life.