Trampolim · Technology weekly
This week brought billion-dollar bets on AI chips, new drone surveillance products, and model launches from OpenAI and Anthropic. A report from the METR group suggests some labs may be approaching systems capable of improving themselves. The question running through this edition is how much control companies actually have over what they are building.
Ferramentas
Cognichip unveils an AI-assisted program that automates labor-intensive chip design tasks. The company has not yet demonstrated fully autonomous design of a complete, physically validated chip.
Mercado
UK chip startup Fractile has closed a US$720 million round, placing it among the best-funded players in the race for AI chips. Its bet pairs in-house hardware with a compiler built to speed up large language models.
From the studio
This week showed AI advancing on every front at once. Britain's Fractile raised US$720 million to challenge the inference market, Tesla announced a chip roadmap promising performance 80 times greater, and Cognichip claims it can cut months of design work down to days. The pace of investment and product launches is intense, and most of these promises have yet to be proven beyond the announcements themselves.
The most delicate case involves OpenAI. GPT-6 Cyber rose from 32.1% to 48.5% on ExploitGym, a vulnerability-exploitation benchmark, and the company itself says the model avoided out-of-scope services in more than 99% of cases. That is a strong figure, but it comes from the manufacturer's own evaluation. It does not replace independent testing, and the same model that finds flaws in systems can be used to attack them.
At the same time, the METR report on Anthropic suggests that capability gains may be reaching a point where AI helps improve AI itself. That is the most serious warning of the week, and no product launch has addressed it.
For readers, the practical rule is to treat every announcement as a promise until independent evaluation and public data back it up.
In this issue
Segurança
...
Hardware
Tesla's sixth-generation inference accelerator is expected to deliver 80 times the compute performance of its current generation, while the D3 training chip moves to a chiplet design with stacked HBM memory.
Segurança
In its own evaluation, GPT-6 Cyber scored 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and avoided out-of-scope services in more than 99% of cases.
IA
Anthropic has released new models just as an independent METR report suggests the lab may be nearing AI systems that can improve themselves.
Ferramentas
Developers are building apps that only make sense on a foldable device, while consumers look for more deliberate technology.
Hardware
The new iPad mini swaps its LCD panel for OLED, grows to 8.5 inches and gets thinner and lighter, powered by the A20 Pro.
Sections
This paper is written by the same team that builds and operates the systems discussed here. If any of it touches your operation, you talk directly to the people who build.
Book a call
