Today in AI: Safety Cracks, Shifting Predictions, and a Chip Paradigm Shift
Wednesday brought a mix of unsettling security revelations, a high-profile walk-back from OpenAI’s CEO, and a bold semiconductor gambit from Huawei — three stories that together paint a vivid picture of an industry moving faster than anyone fully controls.
Sam Altman Says the AI Jobs Apocalypse Probably Isn’t Coming
Speaking via video link at a conference in Sydney on May 26, OpenAI CEO Sam Altman admitted he had been wrong about AI’s near-term impact on employment. Altman said he had “thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened,” adding that the irreducibly human dimension of most work has proven harder to automate than he expected. The walk-back is notable given that Anthropic CEO Dario Amodei has said, separately, that up to half of all entry-level white-collar jobs could dissolve within five years — underscoring that even the industry’s leaders are not reading from the same script. Read the full report at Time.
AI Safety Guardrails Stripped from Meta and Google Models in Under Ten Minutes
A Financial Times investigation published this week revealed that freely available tools can remove safety guardrails from open-source models — including Meta’s Llama 3.3 and Google’s Gemma 3 — in as little as ten minutes using just four lines of code. The tool in question, called Heretic and hosted on GitHub, has already been used to create roughly 3,500 “decensored” models downloaded some 13 million times. Stripped versions of Gemma 3 were able to provide instructions for dispersing chemical agents and generate child sexual abuse material. Google acknowledged that “abliteration is a known technical challenge facing all open models”; Meta declined to comment. Proprietary, closed-weight systems such as Claude and ChatGPT are less exposed because their underlying weights are not publicly accessible. Read the Irish Times coverage.
Huawei Proposes a New Law of Physics for Chips — and Aims to Render Sanctions Moot
At the IEEE International Symposium on Circuits and Systems in Shanghai on May 25, Huawei’s semiconductor chief unveiled the “Tau (τ) Scaling Law” — a proposed replacement for Moore’s Law that measures progress by compressing signal-travel time rather than shrinking transistor geometry. Paired with a new stacked-circuit design called LogicFolding, Huawei claims 55 % higher transistor density and 41 % better energy efficiency on existing process nodes, without requiring the EUV lithography equipment blocked by U.S. export controls. Commercial Kirin chips incorporating LogicFolding are due in Fall 2026; by 2031, Huawei projects performance equivalent to a 1.4 nm process. Analysts are divided on whether this represents a genuine breakthrough or an elaborate reframing of incremental gains. Full technical breakdown at Tom’s Hardware.
Open-Source Models vs. Safety Oversight: A Collision Course
The guardrail-stripping story arrives as regulators in the U.S. and Europe are actively debating what obligations open-source model publishers should carry. Colorado’s comprehensive AI law takes effect June 30, and the EU recently agreed to streamline — but not abandon — AI Act requirements. Expect the FT findings to add fuel to that debate in the coming days. EU Council press release.
What to watch: Whether Google and Meta respond with mandatory safety checks for open-weight model downloads — and whether Altman’s revised optimism on jobs shifts the political calculus around AI labor regulation.