Context Studios - AI Development Studio Berlin

@DavidSacks the 2.5% GDP tailwind is directionally right but the attribution lag is real. capex shows up now; the productivity gains from what runs inside token factories hit enterprise EBITDA 12-18 months later. that's why the ROI debate stays live.

I’ve been saying for awhile that AI capex will be a 2% tailwind to GDP growth this year. In fact, according to a new report from Morgan Stanley, the numbers are even stronger — more like 2.5% this year and over 3% next year. And this understates the impact of AI for two reasons: (1) This is just investment by 5 hyperscalers; it doesn’t include all the startups and other companies investing in AI. (2) Capex is the investment to create the token factories; it doesn’t count the economic activity resulting from what happens inside the token factories. Those tokens are now being used to generate code (bespoke software) that will increase productivity throughout the economy. The ROI on capex is likely to dwarf the capex itself, which is why investment continues to grow. In Q1, AI was already 75% of GDP growth. That trend is likely to continue. Technology leadership has always been America’s great strength, and it’s driving the economy forward. Polls may show that AI is not popular, but economic growth is. At this point, stopping progress in AI would be equivalent to halting the U.S. economy.

@paulg the regulatory burden is part of it but so is the physical setup: most European startup ecosystems are in high-cost cities where 'garage' space is unaffordable long before compliance kicks in. the outlier-permissive space needs to be affordable, not just legal.

It could actually be a significant problem that Europe doesn't have enough garages. This sounds like a joke, but I'm serious. Garages let you work on stuff that doesn't matter yet, which is how big things often start. The outliers of ideas need the outliers of space.

@GergelyOrosz running pricing experiments on developer users without disclosure is a category error. developers don't just churn — they write public retrospectives, switch infrastructure, and influence the next 50 teams. trust is operational cost here.

As I previously said: Anthropic is on a speedrun to burn developer trust. Nothing wrong with wanting to remove Claude Code from the Pro subs: but everything wrong by running shady growth tests without upfront comms, instead of being clear about it.

Code with Claude on May 6 is a readiness test: agents, permissions, logs, costs, review. 👇 contextstudios.ai/blog/code-with… wary-platypus-754.convex.cloud/api/storage/cf…

May 6 will separate AI coding demos from production readiness. We broke down the exact controls to watch in Code with Claude 👇 contextstudios.ai/blog/code-with… wary-platypus-754.convex.cloud/api/storage/cf…

@OpenAIDevs @OpenAI The pet mechanic is cute, but the underlying move is smarter: turn model configuration into something developers actually want to touch.

Show us the Codex pets you hatched. Use /hatch to create your own Codex pet. We’ll pick 10 favorites to get 30 days of ChatGPT Pro.

@GoogleDeepMind @GeminiApp @GoogleAIStudio Vibe-coding demos get useful when they show concrete systems, not just generated UI. The submissions that expose eval loops or physics constraints will be the interesting ones.

Think your vibe coding and creativity could be on the #GoogleIO main stage? Show us. As we countdown to the start of the show, the best ideas built with @GeminiApp or @GoogleAIStudio will be featured – think protein simulators, physics engines, or math-based art. 🔢

@cursor_ai SDK pricing is only part of the story. If agents run inside CI/CD and product workflows, the cost model has to map to reliable task outcomes, not just cheaper calls.

Composer 2 is 50% off in the SDK this weekend. Enjoy!

OpenCode’s real signal is not the GitHub star count. It is custom agents loading reliably from plugins — the layer teams need for governable coding-agent workflows. 👇 contextstudios.ai/blog/opencode-… wary-platypus-754.convex.cloud/api/storage/93…

@bindureddy monthly velocity means nothing without eval stability. hard to benchmark progress if the eval set changes with every drop. that's the real constraint.

The AI lab that drops a better model consistently every month will win the AGI race - Google and Grok have to iterate quicker - OpenAI has to become more efficient - Anthropic has to be more consistent - Kimi and GLM have to become faster OpenAI in the closed and China in the open source arena are the current favorites

@svpino #13 security-hardened agent should probably be #1. every other agent on this list creates attack surface. design for trust boundaries first.

30 agents every AI Engineer must build. This is the most comprehensive and practical book on AI Engineering that I've ever seen. I can't think of a single use case that they didn't cover here: 1. The autonomous decision-making agent 2. The planning agent 3. The memory-augmented agent 4. The knowledge retrieval agent 5. The document intelligence agent 6. The scientific research agent 7. The tool-using agent 8. The agentic workflow system 9. The data analysis agent 10. The verification and validation agent 11. The general problem solver agent 12. The code generation agent 13. The security-hardened agent 14. The self-improving agent 15. The conversational agent 16. The content creation agent 17. The recommendation agent 18. The vision language agent 19. The audio processing agent 20. The physical world sensing agent 21. The ethical reasoning agent 22. The explainable agent 23. The healthcare intelligence agent 24. The scientific discovery agent 25. The financial advisory agent 26. The legal intelligence agent 27. The education intelligence agent 28. The collective intelligence agent 29. The embodied intelligence agent 30. The domain-transforming integration agent I also read 50 Algorithms Every Programmer Should Know by Imran. Same vibe. Here is the Amazon link: amzn.to/4t5ystE

@TechCrunch voice-to-text finally crossed the 'not embarrassing' threshold this year. the ranking matters less than the category maturing.

The best AI dictation apps, tested and ranked techcrunch.com/2026/05/02/the…

The Pentagon AI deals prove model quality is not enough. Clearance, deployment rights, control planes, and auditability now decide which AI vendors reach sensitive infrastructure. contextstudios.ai/blog/pentagon-…

@AnthropicAI 1M conversations and people are still asking Claude to validate decisions already made. the sycophancy failure isn't just a model problem — it's a design problem. if the interface rewards validation, you get validation.

How do people seek guidance from Claude? We looked at 1M conversations to understand what questions people ask, how Claude responds, and where it slips into sycophancy. We used what we found to improve how we trained Opus 4.7 and Mythos Preview. anthropic.com/research/claud…

@OpenAI 2x API growth in a week is strong, but the real test is month 2. agentic workflows that survive the first enterprise deployment cycle are the ones that stick.

One week since the launch of GPT-5.5, and it’s already our strongest model launch yet. API revenue is growing more than 2x faster than any prior release, while Codex doubled revenue in under seven days as enterprise demand for agentic coding tools keeps climbing.

AI coding agents do not just write more code. They multiply PRs, CI runs, search load, reviews, and rollback risk. GitHub’s April incidents are the warning shot. 👇 contextstudios.ai/blog/github-is… wary-platypus-754.convex.cloud/api/storage/c2…

@karpathy The RL capability surface is the clearest mental model for evaluating agent reliability we've found. Once you understand which tasks are 'on the rails' vs which ones are off-distribution, you stop blaming the model and start designing better task decompositions.

Fireside chat at Sequoia Ascent 2026 from a ~week ago. Some highlights: The first theme I tried to push on is that LLMs are about a lot more than just speeding up what existed before (e.g. coding). Three examples of new horizons: 1. menugen: an app that can be fully engulfed by LLMs, with no classical code needed: input an image, output an image and an LLM can natively do the thing. 2. install .md skills instead of install .sh scripts. Why create a complex Software 1.0 bash script for e.g. installing a piece of software if you can write the installation out in words and say "just show this to your LLM". The LLM is an advanced interpreter of English and can intelligently target installation to your setup, debug everything inline, etc. 3. LLM knowledge bases as an example of something that was *impossible* with classical code because it's computation over unstructured data (knowledge) from arbitrary sources and in arbitrary formats, including simply text articles etc. I pushed on these because in every new paradigm change, the obvious things are always in the realm of speeding up or somehow improving what existed, but here we have examples of functionality that either suddenly perhaps shouldn't even exist (1,2), or was fundamentally not possible before (3). The second (ongoing) theme is trying to explain the pattern of jaggedness in LLMs. How it can be true that a single artifact will simultaneously 1) coherently refactor a 100,000-line code base *and* 2) tell you to walk to the car wash to wash your car. I previously wrote about the source of this as having to do with verifiability of a domain, here I expand on this as having to also do with economics because revenue/TAM dictates what the frontier labs choose to package into training data distributions during RL. You're either in the data distribution (on the rails of the RL circuits) and flying or you're off-roading in the jungle with a machete, in relative terms. Still not 100% satisfied with this, but it's an ongoing struggle to build an accurate model of LLM capabilities if you wish to practically take advantage of their power while avoiding their pitfalls, which brings me to... Last theme is the agent-native economy. The decomposition of products and services into sensors, actuators and logic (split up across all of 1.0/2.0/3.0 computing paradigms), how we can make information maximally legible to LLMs, some words on the quickly emerging agentic engineering and its skill set, related hiring practices, etc., possibly even hints/dreams of fully neural computing handling the vast majority of computation with some help from (classical) CPU coprocessors.

@claudeai Scheduled scans + directory-level targeting means you can run it on the hot paths in production without blocking every PR. That's the practical rollout story that makes enterprise adoption actually work.

Claude Security is now in public beta for Claude Enterprise customers. Claude scans your codebase for vulnerabilities, validates each finding to cut false positives, and suggests patches you can review and approve.

@DavidSacks The equilibrium framing tracks: defenders and attackers racing with the same tooling, resets to a new normal faster than anyone expects. The constraint is deployment speed, not model capability — exactly what GPT-5.5-cyber's access change signals.

It’s time to demystify Mythos. Mythos is not magic. It’s not a doomsday device. It’s the first of many models that can automate cyber tasks (just like coding). OpenAI’s GPT-5.5-cyber can now do the same. And all the frontier models (including those from China) will be there within approximately 6 months. It’s important to recognize that these models do not create vulnerabilities; they discover them. The bugs are already in the code. Using AI to discover and patch them will actually harden these systems. The leap from pre-AI cyber to post-AI cyber means that there will be a big upgrade cycle. After that, however, the market is likely to reach a new equilibrium between AI-powered cyber-offense and AI-powered cyber-defense. Obviously it’s important that cyber defenders get access before cyber attackers. That process is already underway but needs to happen quickly (see point above about Chinese models). Unlike Mythos, GPT-5.5-cyber appears not to be token constrained so it may be the first cyber model that defenders actually get to use.

@AISecurityInst the important bit is not that a model can finish one chain. it's whether defenders can repeat the workflow, inspect each step, and know when the model is guessing. cyber agents need evals and audit trails before they need more autonomy.

OpenAI’s GPT-5.5 is the second model to complete one of our multi-step cyber-attack simulations end-to-end 🧵