Oct 12 · AI Daily

An official billing adjustment has been announced: when invoking GPT-6.1 Sol from subscription allowances, long-context fees will no longer be charged.

Is it true?

Google will eliminate LeetCode-style coding interviews, replacing them with an assessment of how candidates use AI tools.

Keep the main point in mind: use Gemini!

OpenClaw has released version v2026.10.1, adding installed skill discovery, MCP Apps beside the chat, support for backing up to external storage or Cloudflare R2, and a simplified Mac installation process. This release includes numerous bug fixes, combining a total of 2,425 PRs from 402 contributors. See https://docs.openclaw.ai/releases/2026.10.1 for details.

For those with less-than-stellar reading skills who still want to learn how to make other people's stuff work well too

(Quoting tweet: OpenClaw v2026.10.1, here is what you need to know:)

Seriously, this actually works!

Think of Grok @Bot like hiring a super smart, 24/7, extremely resourceful employee.

You can easily use it from your phone in minutes:

https://apps.apple.com/app/id6794501026

Lauren Tan suggests that as agents write code on a massive scale, traditional principles prioritizing maintainability and clean code need an update, and the meaning of good code should be redefined for 2026. She points out that clean code historically came at the cost of performance and stems from an era of humans maintaining code, and new principles for agent-friendly architectures may develop in the future. Discussions she previously cited also suggested that good code should be correct, bug-free, run fast, and be low-cost, and argued that productivity should differ noticeably once agents are used well.

According to the Wall Street Journal citing people familiar with the matter, Amazon is in talks to acquire San Francisco AI startup Decart at a potential valuation of around $7 billion. Decart focuses on developing cross-chip adaptation software to help developers switch between different AI chips, and has previously raised over $400 million in cumulative funding, with a valuation approaching $4 billion in a May funding round. Previously, Anthropic had discussed an acquisition for about $6 billion, which reports in September indicated had been dropped.

The Verge points out in an analysis that following Donald Trump's announcement renaming AI to super intelligence—and stating that those who continue to use the term artificial intelligence are enemies—several Silicon Valley tech giants have followed suit.

Apple will release a smart home hub named HomeView on Tuesday, featuring the A18 chip and Siri, equipped with a display to control other smart home devices, offered in both countertop and wall-mounted models. On the same day, Apple will also update the Apple TV set-top box and HomePod mini, and plans to announce a partnership with LG for smart locks, thermostats, and doorbells.

TechCrunch Disrupt 2026 will take place October 13–15 at Moscone West in San Francisco, with over 10,000 founders and investors attending. More than 250 speakers will take the stage across 200+ sessions, including Cat de Jong from Anthropic, Les Karpas from NVIDIA, and Andrew Feldman from Cerebras Systems.

UK Minister for AI Kanishka Narayan will announce enhanced cybersecurity measures and call on frontier AI R&D companies to take more steps to safeguard system security, emphasizing that the UK cannot rely entirely on tech giants for "self-regulation" of models and must build more independent, controllable AI infrastructure. The UK AI Safety Institute (AISI) has partnered with OpenAI, Google DeepMind, and Anthropic to test over 30 models, with its findings prompting actual changes prior to model deployment.

Tibo announced that the default context length for GPT-6.1 Sol in Codex and Work has been increased from 272k to 500k, an increase of about 80%. This move reduces the occurrence of compaction and helps tasks that require long-term retention of large contexts and details. Users can still manually override the setting, with a maximum model context length of 1M.

Quoting Karpathy's analysis, the author suggests that the global population of roughly 8 billion is divided into four tiers based on their level of exposure to large model capabilities. About 6 billion people have almost no contact with large models, 1.1 to 2 billion occasionally use free products, about 20 million heavily use frontier models and agents, and roughly 5,000 have access to internal AI lab systems, with this cognitive gap continuing to widen. The author adds that Chinese users have a tier 3.5 regarding whether they can use Claude Code.

Raise Financial Services is partnering with Cartesia to integrate voice models into its proprietary real-time voice infrastructure, providing multilingual voice experiences for Indian financial markets. Cartesia's Sonic text-to-speech model supports 11 Indian languages and code-switching within a single generation, and is currently handling conversational interactions like KYC in Dhan's voice bot, where over 75% of interactions use non-English languages, with Hindi accounting for over half.

Ethan Mollick reposted Andrej Karpathy's tweet, in which Karpathy estimates that roughly 75% of people have almost no exposure to LLMs, about 20% are light users of free tiers, about 20 million math and coding professionals get to see agents complete projects in weeks that used to take months, and roughly 5,000 internal lab staff observe thousands of agents collaborating over weeks on large software projects, vulnerability discovery, cybersecurity offense and defense, scientific discovery, and frontier math progress. Karpathy suggests this cognitive funnel is driven by problem ambition, jagged capabilities, and access tiers. Mollick adds that even though many of these 5,000 internal staff have not fully utilized advanced model capabilities, their applications are concentrated in cybersecurity, mathematics, and AI R&D, while many other fields and problems remain to be advanced.

The vLLM Semantic Router team has released the Decision 3.0 series of open-source decision models, which read text, JSON, and images, returning probabilities rather than generating text, all under the Apache-2.0 license.

Sakana AI has released Multi-Agent Self-Supervision (MASS), which uses a shared model through virtual sub-agent teams to solve tasks, search for better multi-agent workflows, and then uses its own judgment to select the best results and distill them back into the model, achieving recursive self-improvement.

Karpathy estimates that about 6 billion people globally (around 75%) have virtually no exposure to LLMs, roughly 1–2 billion people occasionally use free-tier ChatGPT-like products, about 20 million people directly use frontier models in mathematics and coding, and only about 5,000 people have access to frontier-grade internal systems. He suggests the funnel is driven by problem scale, jagged distributions of system capabilities, and access tiers, meaning different groups perceive the exact same set of systems very differently.

Andrej Karpathy reposted and expanded on his April views, arguing that the gap in outside consensus regarding LLM capabilities is turning from "two groups talking past each other" into a sharply narrowing funnel.

Following OpenAI's release of hundreds of AI-generated papers targeting open math problems all at once, some Northwestern University students began questioning whether they should remain in the field of mathematics. Professor Antonio Auffinger stated that his current priority is retaining talent and helping them grow; Reihel, who was quoted, noted that senior mathematicians jumping in to do student thesis topics violates academic norms, and that AI is stealing training topics that used to take graduate students years to complete.