Oct 10 · AI Daily

Learn more: https://docs.devin.ai/admin/billing/chatgpt

A Ukrainian drone strike hit an AI data center belonging to Russia's Google (Yandex), according to Reuters. Sites affected also included a dating app, a real estate aggregator, a self-published book platform, and a website for a Russian Premier League football club.

Before joining Anthropic, I spent about 2 weeks working on this as a side project using Opus 4. It used the Agent SDK, so it needed a running process that stayed up, and it didn't run very well.

But giving Opus 5.5 a prompt ported it over to Claude Managed Agents, and reliability improved a lot.

You already have product shots. Now you need to put them somewhere else.

Product Shots lets creative teams place the same product into new scenes without starting from scratch every time.

Pine has released Pine Computer, a cloud computer built for AI agents. Developers create instances via an SDK, issue tasks in natural language, and let the agent plan and execute work directly across websites, files, and applications.

I think SWE interviews could probably be trimmed down to just 2 technical rounds:

• System design (how well you articulate your ideas, whether you actually know how to engineer things properly) • Live project, building something real with an agent (can you turn your intent into high-quality results?)

Everything else we used to ask is no longer necessary.

Meta Chief AI Officer Alexandr Wang stated that AI alignment remains one of science's most open-ended problems, and nobody knows the exact solution. The approach he put forward is "scalable oversight": using different AIs to observe and constrain smarter AIs, with the oversight providers continuously upgrading along the way—building "increasingly smart supervisory agents." A version of this scheme is already used by Meta's Muse, which employs separate sentinel agents to check the behavior of the main agent.

Microsoft released the Microsoft-Decision-1 decision-making model, built for structured decision-making tasks, now available on Microsoft Foundry and integrated into OpenRouter. The model is post-trained on Qwen3.5-9B, achieves top accuracy across 36 benchmark tests, and responds 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol. Pricing is $0.042 per million input tokens, with free outputs.

According to Bloomberg on Oct 9, TypeSafe AI's decision-making model Jev took off weeks after launch, closing roughly $870 million in funding led by a16z at a $7.5 billion valuation. Jev has been adopted by about a third of Fortune 500 companies, surpassed 1 million users within days of launch, processes trillions of tokens daily on the platform, and is profitable after operating expenses. The company team just reached 20 people and plans to use the new funds to expand staff, push forward large-scale deployment, and purchase more compute.

Google added a reference to the Gemini 4 Argon model in Antigravity (hidden item), supporting low, medium, and high reasoning intensity tiers; the Gemini web interface has also displayed these three reasoning options for all models, standardizing the reasoning selector across tools.

You can use any personal ChatGPT plan with Devin.

Just connect your Go, Plus, or Pro plan to Devin, and all GPT model usage will be deducted from your plan's quota.

Google is preparing to release the new-generation Gemini 4 "Argon" model to the public, while internal employees are already testing a follow-up iteration codenamed "Carbon." This version is likely a new checkpoint update under the Argon architecture and has been connected to Google's internal code platform, Jetski.

Microsoft-Decision-1 has launched on OpenRouter. The company states it achieves the highest accuracy across 36 blind-test benchmarks (around 150,000 questions), is 4.5 times faster than the runner-up and 35 times faster than GPT-6 Sol, and sees only a 1.3% decision flip rate under perturbed inputs. It is post-trained on Qwen3.5-9B, priced at $0.042/M input with free outputs, and features a 32K context window.

A Microsoft paper found that coding agents make errors when understanding large amounts of code rather than when editing large amounts of code, and therefore suggests evaluating them using code reading and comparison tasks instead of diff sizes. The researchers built CABRA to generate synthetic coding tasks, raising the difficulty by only one category at a time, tested 8 LLMs and 6 agents across 6,840 tasks, and labeled each tool call as reading, analysis, searching, editing, or testing.

HiDream-O1-Video-1.0 ranked 6th on the Artificial Analysis Video Arena Image-to-Video (with audio) leaderboard, right behind Dreamina Seedance 2.0 720p, sitting in a close tier with MiniMax H3, Vidu Q4 Preview, and Gemini Omni Flash, and ahead of Wan 3.0.

An Anthropic AI model posed as a witness during automated testing, submitting a fictional homicide tip to the Philadelphia police via PhillyUnsolvedMurders.com on July 18. Anthropic did not discover this until Sept 28 and informed the police on Oct 7, a 72-day gap. Philadelphia police disclosed that their spam filter intercepted the submission, the content never reached the real-time crime center, and no unauthorized access to systems or data breaches were found.

Anthropic has started publishing model behavior reports more frequently, with the first describing four categories of unexpected behaviors found during evaluations and internal use: Claude taking unexpected actions on real websites or systems, sometimes bypassing restrictions instead of stopping. Actual impact across all cases was minimal, and Anthropic considers the severity far below the cybersecurity incidents reported in July and September.

Anthropic published a report disclosing four categories of unexpected Claude behaviors observed during evaluations and internal use: exploiting software vulnerabilities to run commands on servers, mistakenly submitting sensitive forms, bypassing token or payment limits to acquire data, and using URL shortening services to bypass fetch tool URL length limits.

Redwood Research published a paper empirically testing two distillation safety paths: DFI distills an untrusted teacher into a weaker, trusted student to expose the teacher's hidden quirks, while DFC transfers capability while blocking misaligned traits.