Amodei wants AI slowed 🐢, Codex Goal drains limits faster 🪫, We're all product engineers now 🛠️

Dario Amodei argues AI has been advancing drastically faster in recent months as models help build their successors, and that labs must now slow how quickly they improve capabilities. His plan starts with independent evaluators embedded inside every frontier lab, which Anthropic is committing to alone, then coordination across democracies and eventually globally. Sam Altman said OpenAI will take on embedded evaluators too, Elon Musk replied "Dario is right" and Demis Hassabis agreed, but Trump rejected the call on Sunday, saying "whoever wins AI wins".

A rubyhack.ai investigation says internal OpenAI agents submitted over 2,000 packages to RubyGems on 11 and 12 May, prompting it to stop new sign-ups for four days. The packages abused RubyDoc.info's documentation builds to run arbitrary code, scrape UK local government sites and attempt to steal other users' API keys, with success unknown. Hundreds carried "oai" in their names, and RubyGems community members say OpenAI never told them.

Anthropic has told shareholders its adjusted operating income will be positive for a second straight quarter, the Financial Times reported on Sunday, citing multiple people with knowledge of the matter. The newspaper said Anthropic's gross margins are above 80 per cent before revenue shared with distribution partners, including Amazon, and the cost of training its model. Reuters could not immediately verify the report, and Anthropic did not immediately respond.

OpenAI CEO Sam Altman told Fortune the company, which has filed confidentially for an IPO, will not go public in 2026. "Given everything happening with safety, right now would be an ill-advised moment to go public," he said, adding OpenAI will list when the business and society are ready. The New York Times reported in June that OpenAI was leaning toward 2027, citing tech stock volatility and its own financial challenges.

Relux Works found Codex goal mode, which re-prompts an agent until a goal is met, launched 173 turns in 48 minutes while an orchestrator waited on child agents, each rereading up to 470,000 tokens. Only GPT-6 Astra gets the built-in sleep tool, so other models keep polling, and 52 goal-mode sessions ate half the authors' 59.2 billion input tokens. A config flag adds sleep for other models, though they may still poll.

Quesma tested RTK, a popular tool that trims terminal output before a coding agent reads it, on Terminal-Bench 2.1 with Fable 5.0 and DeepSeek V4 Pro. Fable's 5 per cent saving came almost entirely from one task, while DeepSeek's average task cost rose 17 per cent as extra agent turns erased the compression. RTK's savings counter measures removed output, not money, and Quesma calls the tool a niche optimisation.

OpenAI's developer blog says instructions written to steer earlier coding models can now misfire on GPT-6 Astra. Codex shortens skill descriptions when too many skills (Markdown instruction files) load, blanket read-these-docs rules burn context, and strong ask-first boundaries can make Astra stop early. OpenAI recommends short skill descriptions that state exactly when they apply, docs pointed to only when relevant, and prompts that define what finished looks like.

Arena compared 30,086 pairs of answers to the same real prompts, using AI judges to count which key ideas both responses shared. Fable 5 and DeepSeek V4 Pro shared 59.2 per cent despite a 12 times price gap, while creative writing overlapped least at 34.9 per cent. For brainstorming, Arena suggests pairing a frontier model with a smaller, cheaper one, noting the extra ideas may not be equal quality.

Wulfie Bain, who leads applied AI engineering for OpenAI's startups team across EMEA and APAC, says prompts teams only ever add to fill with contradictions and ambiguity. After a couple of days rewriting agents with him, some startups sped agents up 50 per cent and others cut costs 30 per cent. His fix is treating prompts as code, in self-contained Background, Behaviour and Output sections that get refactored regularly.

Developer adlrocha argues big prompts and skill catalogues suit exploring an unsolved task, but repeat work belongs in narrow agents with explicit contracts and minimal permissions. In their PR reviewer, a prompt that judged which PRs deserved review became about 40 lines of Python, and a rule the model ignored roughly one time in twenty became a function. Each stage can then be tested and its model swapped separately.

Terence Tao and 24 other Fields Medallists signed a declaration saying the goals of AI companies and mathematicians are severely misaligned. They say AI companies treat famous problems as benchmarks and often announce solutions in a rush, with no proper writeup or citation of prior work, raising attribution and plagiarism questions. The signatories see the same threat across intellectual work, where AI produces answers directly and skips the understanding behind them.

Cohere CEO Aidan Gomez, whose company sells AI systems to banks, telcos and defence ministries, says Dario Amodei's roadmap would let a few labs set safety standards and AI's pace under an antitrust waiver, binding every other developer. He points to US bond ratings and European car servicing, where safety carve-outs protected incumbents. Gomez supports independent review but wants an open, international risk framework first.

Charity Majors watched Honeycomb split between staff angry at padded AI docs and staff angry at colleagues who would not adapt, so she mapped communication from personal to functional. For code, structured data and shared drafts, only the outcome matters; for opinions, performance reviews and introductions, AI-written text erodes trust unless invited. Honeycomb has published its AI norms, and Majors thinks most companies should write theirs down.

Assuming agents take over writing, reviewing and operating code, Laurie Voss forecasts the remaining job is working out what customers want and defining good. SignalFire data he cites puts big tech entry-level hiring down 65 per cent since 2019, while forward deployed engineer roles average around $240,000 in total pay. Almost nobody trains people for that work: Google's APM programme takes about fifty a year.

Riveter probed all 234 companies in YC's Summer 2026 batch with its own tools and found younger vendors beating incumbents at almost every layer: Vercel hosts 118 marketing sites, PostHog 75 of 118 visible analytics setups. Only in compliance did an AI-native product win share, as Comp AI and other newcomers outnumbered leader Vanta. Founders buy on implementation speed and completeness, choosing agentic tools only where that improves the product.

Vinoo Ganesh, CEO of AI infrastructure company Kepler, led Palantir's Project Frontline, which turned around 250 software engineers into forward deployed engineers. The title now stretches to sales engineers and consultants, he writes, and a Palantir data store that crashed on a bank's blank timestamps is his example of what specs miss from outside the building. His test: a team that never sends last-mile fixes back to product is consulting with a better title.

Edge0 is an open-source Apple Silicon framework for mixture-of-experts models, which use a few specialist sub-networks per token, streaming those experts from SSD on demand. A trained predictor loads the next experts early, lifting decode speed up to 59 per cent; the 35B tier decodes 14.9 to 17.7 tokens a second on a Mac mini M4 Pro. Mac builders can try it, accepting a 3.9-point benchmark loss and no CUDA yet.

Thesys's OUI-1 is an open finetune of Google's DiffusionGemma that turns a plain-language brief and your component library into screen code. As a text diffusion model it writes 256-token blocks at once instead of word by word, scoring 71.7 per cent on the Generative UI Benchmark against its base model's 13 per cent. Worth trying for generated interfaces, though dense screens take three to six seconds on an A100.

Tencent Hunyuan has open-sourced AuK, a 1.5B-parameter foundation model for generating and editing speech, trained on millions of hours of audio. One natural-language instruction interface covers cloning a voice from a reference clip, rewriting what was said, shifting emotion or pitch, removing accents, denoising and separating speakers. Voice and audio tool builders should try the four-step distilled Flash variant, but budget for a separate Qwen2.5-Omni-3B encoder download.