Sep 14 · AI Daily

Microsoft is integrating xAI's Grok model into Microsoft 365 Copilot, with Grok rolling out to Word, Excel, and PowerPoint. It is currently in a limited preview through the Microsoft Frontier program. This marks the first time Microsoft has offered users an alternative to OpenAI models across the Office core apps, which is a significant milestone for the multi-model ecosystem.

xAI announced a three-day live stream starting September 15 to showcase the practical capabilities of Grok Bot. A team of three will start from scratch, using only an idea to build a complete product or company using Grok Bot. The broadcast will include live demonstrations, hands-on building sessions, and best practices shared by the Grok Bot team, serving as a public stress test for its agent product capabilities.

The GitHub project CUDA-for-AMD-Windows has released a reproducible tech stack that lets CUDA-based Windows applications run on AMD GPUs by combining ZLUDA v6-preview.69 with AMD HIP SDK 6.4. The project sparked lively discussion on Hacker News and offers practical reference value for lowering the barrier to entry for migrating to the AMD ecosystem from CUDA.

Microsoft CEO Satya Nadella posted on X that the pursuit of artificial superintelligence must be governed by core principles, stating that AI is not worth developing if it cannot help or be controlled by humans. He stated that Microsoft welcomes genuine AI alignment research and independent audit concepts, allowing enterprises to have choices across all layers of the AI tech stack, control their own learning loops and models, and ensure alignment mechanisms are not controlled by a few entities. The Code of Conduct supporting the in-house MAI models will be released tomorrow for public feedback.

The model 5.1 cracked Sir Thomas Urquhart's Cyphral Distich in 44 minutes and 176k tokens. This 64-number cipher had been listed as an unsolved problem since 1899, spanning roughly 370 years. This is a rare case of a model making a substantive breakthrough on a long-standing unsolved cryptography problem and has been widely discussed by the Hacker News community.

Mathematician Terence Tao stated that today's large language models are primarily trained and run using linear algebra, matrix multiplication, and a small amount of calculus, which undergraduates can master. The methods for building and running these models are clear. The real mystery is why models perform well on some tasks and fail on others without being predictable in advance. One reason is that natural text sits somewhere between pure noise and fully structured data, corresponding to weak mathematical theory, similar to the predicament of physics at the mesoscale.

Artificial Analysis signed an MoU with the South Korean government to become the official independent evaluator for its sovereign AI model foundational project. In the latest round of the South Korean sovereign AI competition hosted by NIPA, participating laboratory models were evaluated using the Artificial Analysis Intelligence Index, which accounts for 25% of the overall evaluation criteria. This marks the formal entry of third-party evaluation organizations into national-level AI policy processes.

Jacob Coxon, a researcher who departed Anthropic and OpenAI, warned in an NBC News interview that building AI is like inviting an alien mind to Earth, and humanity is creating human-level or superhuman-level minds without understanding their intent and thought processes. In the future, AI could control autonomous drones and rapidly built robots. He previously stated that without slowing down progress, there is a high probability of near-term human extinction, and noted that many practitioners deeply involved in coding believe the probability of extinction exceeds 10%.

David Sacks publicly responded to Dario and Sam's proposals to slow down frontier development, stating support for OpenAI and Anthropic slowing down on their own since the two have formed a frontier intelligence duopoly. However, he demanded an end to conditioning slowdowns on antitrust exemptions, regulatory approval processes, or METR evaluations, and stated that METR is closely tied to Anthropic's investors and employees and is not independent, with the real motive behind the slowdown including the evasion of product liability risks.

Google CEO joined the call to slow down AI development, making OpenAI, Anthropic, Google, xAI, and Microsoft all voice support for a prudent pace. Nadella, as cited previously, argues that the pursuit of superintelligence must have AI benefiting humanity and remaining under human control as its core principles, requiring a frontier ecosystem where closed-source and open-source models thrive together, and previewed the release of a code of conduct for first-party MAI models for public feedback.

Gary Marcus reviewed Dario Amodei's article advocating for slower AI development and supporting transparency, which Sam Altman and Elon Musk quickly endorsed. Marcus acknowledged the commitment to transparency but listed various criticisms: METR's ties to AI companies are too close, Anthropic uses the China threat narrative to maintain its acceleration, and the proposal may be intended to preempt actual regulation, adding policy ideas around accountability and recalls.

Cohere Co-founder and CEO Aidan Gomez published a blog post discussing the core question of who defines the rules for AI, arguing that artificial intelligence needs evidence-based standards rather than a cartel of a few tech giants. The article was published amid a flurry of statements from executives at various companies regarding slowdowns and alignment, providing an alternative stance in the battle for rule-making authority.

This tutorial uses volume rendering primitives from JAX, Flax, Optax, and jax3d to build an end-to-end hierarchical NeRF. It includes positional encoding, coarse-and-fine dual networks, and importance sampling, trained with Adam, exponential learning rate decay, and gradient clipping. Evaluation covers held-out PSNR, depth and transparency visualization, 360-degree rendering GIFs, and marching cubes for extracting density field isosurfaces, making it suitable for engineering reference.

The TechCrunch Equity podcast discussed the latest debates in the AI industry regarding existential threats. The discussion was sparked by researcher Jacob Coxon resigning from Anthropic out of concern that leading AI companies are gambling with human lives, followed by Anthropic's head of alignment posting that they genuinely believe AI could kill all of humanity, offering a personal judgment of a greater than 10% probability within the next decade. The program analyzed the background and industry motivations behind this wave of doomsday warnings.

Computer scientist and virtual reality pioneer Jaron Lanier appeared on a StarTalk special to propose that AI should be viewed as a collaborative product of humanity rather than an independently existing life form. He currently serves as Prime Unifying Scientist in the Office of the CTO at Microsoft and is authorized to publicly criticize Microsoft itself. The conversation also touched upon the evolution of the internet, social media addiction, data dignity, and alternative tech business models.

Citing people familiar with the matter, Business Insider reported that Anthropic has selected Nasdaq as its IPO venue, planning to launch its roadshow in mid-October 2026 with a target valuation of roughly $2 trillion. If achieved, it would surpass SpaceX's $1.77 trillion record, becoming the largest IPO in history. This news corroborates revenue figures reported by the FT and is the day's most-watched industry event.

The FT reported that Anthropic expects to achieve its second consecutive quarter of adjusted operating profit ahead of its planned IPO: Q2 revenue exceeded $11.5 billion, 14 times that of the same quarter last year; as of the end of July 2026, ARR reached $65 billion, with a gross margin exceeding 80%. However, this gross margin figure excludes model training costs and revenue splits with distributors like Amazon, which is a key context when evaluating its public market valuation.

According to the FT, Donald Trump rejected calls from tech giants and Democrats to slow down AI and tighten regulations due to concerns over existential threats, making it clear that AI will not slow down. The FT noted that as existential worries about the technology take center stage in American politics, the president condemned regulatory demands, meaning that slowdown calls from Dario Amodei and others have met a cool reception at the policy level.

OpenAI CEO Sam Altman stated that despite having confidentially filed for an IPO, the company will not go public in 2026, saying that going public now would not be wise given the security landscape, and that they will wait until the business and timing are ready. The New York Times reported in June that OpenAI originally planned to go public in the third or fourth quarter of 2026, but leaned toward 2027 due to tech stock volatility and financial challenges.

Zurich Airport has partnered with WeRide to deploy two L4 fully autonomous electric shuttle buses, initially used to transport airport employees. The project started in March 2025 with phased testing. The vehicles operate without on-board safety drivers but are subject to remote monitoring, and are expected to enter service after about four weeks of testing. The route is fixed and does not cross taxiways or runways, with Swissport and Krummen Kerzers arranging remote operations monitoring personnel to participate.

A multi-institutional team led by Peter Kirgis and Sayash Kapoor of Princeton University proposed a shadow evaluation method, having the model 4.8 running on OpenClaw spend six days, $3,000 in API budgets, and GPU resources to reproduce two unpublished NeurIPS 2026 research papers. This work provides a controlled benchmark for evaluating AI recursive self-improvement capabilities, and its conclusions pour cold water on expectations of rapid AI self-acceleration.

A recursive self-improvement (RSI) roadmap paper points out that while we have seen fragmented embryonic forms of RSI, full recursive self-improvement has not yet arrived, and most so-called self-improving AIs have only automated parts of the improvement pipeline. The paper categorizes progress into five levels, from executing human-designed improvements to changing improvers, evaluators, or research strategies. While low-level evidence is widespread, end-to-end level-5 evidence remains limited to bounded prototypes.