Co-Founder @typewise_app
Zurich, Switzerland
Joined January 2010
- Tweets651
- Following37
- Followers124
- Likes104
Anthropic tested AI agent teams across 12 simulated tasks. On one lending task, a single agent scored 0.1 business/1.0 ethics. The team nearly inverted it: 0.8/0.35. Testing each agent separately does not certify the organization.
1
13
METR says Anthropic's progress on well-specified AI tasks accelerated 1.9x (95% CI: 1.5-2.8x), with high uncertainty. Anthropic redacted a majority of the evidence METR found most informative. Independent access is not public transparency.
13
Canva is changing 22% of its team to get back to startup speed. 1,271 people face role changes, relocation or departure. CEO Melanie Perkins says narrow ownership and handoffs made bigger problems harder to solve. Org debt is paid by people.
1
20
Salesforce agreed to acquire Cresta. The Information reports a $2.7B price. The real asset isn't another general model: it's the loop from customer conversation to CRM action to outcome. In AI customer service, platform control is becoming the moat. Buyers need a real exit path.
Anthropic says internal Claude use at list prices grew 250x from Jan to Sep 2026. In April, that value first passed monthly personnel expense excluding stock comp. But list price is not cost, and usage is not output. AI ROI needs 3 lines: cost, priced value and verified work.
Mistral Large 4 gained 13 points on Terminal-Bench 4.0 in three weeks while using 30% fewer output tokens, then Mistral released the weights. European teams can now benchmark AI sovereignty as an engineering option, not a policy slogan.
More than 40% of GitHub pull requests now start with an AI agent, up from 17% in March. GitHub counted 110M agent-started PRs in September, about 6x March. The bottleneck is moving from writing code to verifying what belongs.
9
Anthropic says an internal AI model found five attack chains in widely used software. One flaw could crash several million internet-facing Windows servers with a few packets. The security unit is no longer one bug. It's the path through several defenses that fail together.
18
Anthropic's S-1 says cloud marketplaces handled 48% of H1 2026 revenue, up from 11% in 2023, and 58% of net receivables. AWS, Google and Microsoft can be suppliers, distributors, investors and competitors. Fast distribution can become someone else's leverage.
154
SpaceX raised $40B in debt to buy GPUs. Musk says its edge is revenue beyond AI data centers, while 200+ neoclouds may consolidate into a few survivors. AI infrastructure is becoming a balance-sheet contest. Credit risk belongs beside utilization and power.
7
US banks are blocking third-party AI agents from customer accounts. A password proves access, not delegated intent. Every enterprise AI agent needs its own identity, scoped permissions and audit trail. Shared logins are not an authorization system.
5
1
42
OpenAI says ChatGPT Ads hit $1B annualized revenue in under 200 days, with 100K+ advertisers. The smartest design choice: ads run separately from answers, and "no ad" is a successful outcome. A revenue system should never get a vote on the answer.
1
22
Reuters reports $35B+ in Nvidia RTX PRO 5500 orders from ByteDance, Alibaba and Tencent. Partners may receive only 5%-10% of requested volume. In regulated AI markets, a purchase order is not a capacity plan.
20
Meta cut a 3-to-4-day engineering task to under an hour. Its 1,300-line orchestrator routes 20+ workflows to 14 specialist agents and does zero domain work. The architecture lesson: keep routing deterministic and reasoning isolated.
8
Anthropic wants to put its mission on the cap table. Subject to stockholder approval, 5% of its shares would go to a public benefit fund, with another 2% possible for every 50% rise in company value. Mission is cheap when it costs nothing. Equity is not.
10
IBM surveyed 3,221 EMEA executives: 80% said an AI incident reached the board or C-suite in the past year, but only 40% changed governance. Escalation is not oversight. Every material incident should end with a recorded decision before agent autonomy expands.
1
14
Sanas says one supervisor can govern up to four AI customer service calls at once, stepping in mid-conversation when needed. That does not remove human judgment. It concentrates it. The hard design problem is routing scarce attention before trust breaks.
7
Microsoft says Maia 300 will deliver 40% more tokens per megawatt than leading GPU systems and up to 7x Maia 200’s rack throughput. The procurement metric is shifting: useful work inside a fixed power envelope, not peak chip performance.
14
Deno is joining Cloudflare to make Workers easier to leave: one binary, one storage bucket, self-hosted. Smart. The strongest platforms win when customers stay even after the exit path becomes credible.
26
