AI News #102: Week Ending September 12, 2025 with 29 Executive Summaries, Top 8 Links, and 1 Lonely Visual

September 13, 2025

About This Week’s Covers

This week’s main cover is a collage of photos from my time outside recently. When I dropped my daughter off at college, I decided to take two months off of my newsletter in order to be present and reset. It’s currently the end of November 2025, even though I’m going back through the weeks I missed.

The rest of the covers were created with my fourteen-week-old GPT rubric + Flux Pro Ultra. I gave GPT-5 a one-sentence description of the theme (vacations and taking a break), and GPT-5 automatically generated 50 cover image prompts and sent them through the Flux Pro API with no supervision.

I’ve included my favorite six covers, below:

This Week By The Numbers

Total Organized Headlines: 304

This Week’s Executive Summaries

This week’s headlines centered around a few key categories. We’ll start with Agents and then go to Open Source, do a little bit about Spatial Understanding and Imagery, then Science, and wrap it up with Business.

Agents

The top story in agentic news is that Google released a dedicated AI mode for their search engine. Rather than AI Overviews, you can go to google.com/ai and the entire interface becomes an AI engine. This is an early preview of where the internet might be headed as page views go away and web searches become like API interfaces. Because it’s so early, I would imagine Google’s AI tool may be a little clunky, but it’s a harbinger of what’s to come.

The rest of Agents news is largely tied to Anthropic and OpenAI.

Anthropic added a web-fetch tool to the Anthropic API. If you give Claude a website’s address, the API can fetch the website and analyze its content.

OpenAI has added full support for Anthropic’s MCP tools protocol within ChatGPT, including things like using GPT to create JIRA tickets or trigger Zapier workflows. Notably, Stripe now supports payments within ChatGPT, using the MCP protocol.

Anthropic released quite a few finance skills this week, particularly the ability to work with Excel files.

You can give Claude an existing spreadsheet to work with or have it generate a new one. According to initial reviews, Claude does a great job with formatting and formulas.

Wharton Business School professor Ethan Mollick shared an example of a robust spreadsheet with 406 formulas (!) generated from a single prompt. After reviewing it, he feels it’s solid.

Another user asked Claude to create a discounted cash-flow financial model, and Claude created a nine-sheet workbook complete with formulas that appears to be viable. The same guy had Claude replicate his profile photo in an Excel file as fun feat of strength.

Adobe announced AI agents across a variety of third-party ecosystems. They appear to have an audience agent that handles recommendations, a customer-journey agent, experimental performance-data agents, data-insights agents, website-optimization agents, and an Adobe product-support agent.

I miss the days when I worked with Adobe’s business products. I now think of Adobe as mostly Photoshop and Premiere; however, for folks in the retail world, Adobe’s acquisition of business tools like Omniture and A/B testing platforms was a thrilling time. If anyone has experience with these Adobe business agents, I’d love to hear from you.

Microsoft launched an audio engine that can read scripts with emotional expressions and perform multiple voices across characters. Most folks see ElevenLabs as the leading voice engine, so it’s interesting to see Microsoft appear on the scene.

Open Source

Ever since DeepSeek showed up in January of 2025 and disrupted the frontier-model landscape by releasing an open-source model that performed at levels only three months behind the pace of OpenAI’s flagship model, there’s been a lot of discussion about open source chasing down closed frontier models… especially with the volume of open models coming out of China.

This week, two models entered the top 10 on text benchmark leaderboards. At number six was the latest Qwen from Alibaba, and Kimi came in at number eight, from a company called Moonshot.

Kimi has been making a lot of headlines, and we’ll have to keep an eye on it over the next few weeks.

Another coding agent called GLM has popped up in the news quite a bit. GLM is from Z.ai. Whenever a model is brand new, I try not to dive too deep into it until it’s had a couple of weeks to settle down from the hype cycle. That said, keep an eye out for Kimi and GLM! I added a category to the newsletter for Kimi’s parent company, Moonshot.

Google’s Gemma 3, another open-source model, is available in the App Store. It can fit on a mobile device and handle speech, text, and image input and output. Google’s Gemma has been one of the strongest small multimodal models for some time, and the idea of on-device multimodal processing is very powerful, with clear use cases for phone features.

Imagery & Spatial Understanding

In image news, Google’s NanoBanana is available in the Gemini API. The potential for both good and bad applications of such a strong image tool is really worth thinking about. On one hand, users can now automate incredibly large volumes of imagery… and of course, that’s exactly the problem.

A user on Twitter this week showed an example using an n8n agent to generate thousands of ad variations with the click of a button.

The idea of even needing to make anything in advance is starting to seem unnecessary. We’re still in an “asset scarcity” mindset of building everything before the user encounters it, but at this point, maybe a year from now, everything will just diffuse in front of us in real time… whether it’s an entire website appearing dynamically or a product image showing us what a sweater would look like in a virtual try-on.

I have a hunch in the coming years, building or creating everything in advance will ultimately feel provincial. I encourage everyone to start exploring the mental shift away from brute-force content generation to on-demand dynamic content.

Andrey Karpathy wrote:

“Chatting” with LLM feels like using an 80s computer terminal. The GUI hasn’t been invented, yet but imo some properties of it can start to be predicted.

1) it will be visual (like GUIs of the past) because vision (pictures, charts, animations, not so much reading) is the 10-lane highway into brain. It’s the highest input information bandwidth and ~1/3 of brain compute is dedicated to it.

2) it will be generative an input-conditional, i.e. the GUI is generated on-demand, specifically for your prompt, and everything is present and reconfigured with the immediate purpose in mind.

3) a little bit more of an open question – the degree of procedural. On one end of the axis you can imagine one big diffusion model dreaming up the entire output canvas. On the other, a page filled with (procedural) React components or so (think: images, charts, animations, diagrams, …). I’d guess a mix, with the latter as the primary skeleton.

But I’m placing my bets now that some fluid, magical, ephemeral, interactive 2D canvas (GUI) written from scratch and just for you is the limit as capability goes to \infty. And I think it has already slowly started (e.g. think: code blocks / highlighting, latex blocks, markdown e.g. bold, italic, lists, tables, even emoji, and maybe more ambitiously the Artifacts tab, with Mermaid charts or fuller apps), though it’s all kind of very early and primitive.

In other imagery news, TikTok owner ByteDance released a strong competitor to NanoBanana. SeedDance 4.0 is good at making high-density infographics and can handle dense text and equations, something historically challenging for image models.

One of my favorite trends is the convergence of image recognition, multimodality, and robotics training and simulations.

Segmentation is the ability to see and identify an object in an image. Depth involves color-coding the object based on how close or far it is in an image. Tracking is following that object across frames.

Sony’s Hawk-Eye system uses dozens of cameras placed around a stadium to build a spatial 4D understanding of the environment in real time. This has clear uses for soccer or ice hockey.

Science

This week we learned about a Facebook group with 89,000 physicians in it. A cardiologist shared a detailed write-up about a patient’s hypertension that he couldn’t diagnose. Another doctor gave the entire write-up to GPT-5 Pro, and in less than 15 minutes, the model figured out what was wrong. The doctor compared GPT-5 Pro’s performance to the best specialist at the Mayo Clinic.

Business

OpenAI and Microsoft announced that they have signed a non-binding memorandum of understanding for the next phase of their partnership.

Oracle signed a $300 billion cloud deal with OpenAI that sent their stock soaring.

Tesla proposed a pay package that could make Elon Musk the world’s first trillionaire.

Perplexity has reportedly raised $200 million at a $20 billion valuation. Perplexity also announced that it is working with the U.S. government to build secure access to premium AIM models.

Robotics company Unitree is targeting a $7 billion valuation in its IPO, planned for Q4 of this year.

Wharton professor Ethan Mollick made an interesting point: SaaS vendors that build their own AI solutions using wrappers are incentivized to use cheaper models and minimal reasoning. This encourages clients to keep outdated prompts or RAG strategies instead of updating as frontier models improve.

And lastly, Microsoft AI CEO Mustafa Suleyman continues to repost a long-form blog from August. In it he urges humanity to build AI to be a tool rather than a proxy for being human. His recurring theme these past few weeks is that we are collectively on the verge of a moment when AI will be able to convince us it is conscious, and he wants to remind everyone that it is not. Reinforcing this message has become almost a daily cadence from him.

This week’s humanities reading is the poem “To Nature” by Samuel Taylor Coleridge.

To Nature
It may indeed be fantasy when I
Essay to draw from all created things
Deep, heartfelt, inward joy that closely clings;
And trace in leaves and flowers that round me lie
Lessons of love and earnest piety.
So let it be; and if the wide world rings
In mock of this belief, it brings
Nor fear, nor grief, nor vain perplexity.
So will I build my altar in the fields,
And the blue sky my fretted dome shall be,
And the sweet fragrance that the wild flower yields
Shall be the incense I will yield to Thee,
Thee only God! and thou shalt not despise
Even me, the priest of this poor sacrifice.

-Samuel Taylor Coleridge

Gettting out into nature is good medicine! Especially once the clocks change and the sun sets.

Full Executive Summaries with Links, Generated by Claude Opus 4.5

Google launches AI Mode for complex conversational search queries
Google has introduced AI Mode, a new search feature that allows users to ask detailed, multi-part questions and receive comprehensive responses powered by its Gemini language model. This represents a shift from traditional keyword search to conversational AI assistance directly within Google’s main search product, potentially changing how billions of users interact with information online. The feature appears designed to compete with ChatGPT and other AI assistants by integrating advanced language capabilities into the world’s dominant search engine.

Google AI Mode – Meet AI Mode Ask detailed questions for better responses google.com/ai https://x.com/OfficialLoganK/status/1964095318254911511

Anthropic adds web browsing capability to Claude API
Anthropic’s API now lets Claude directly fetch and analyze content from any webpage URL without requiring developers to build their own web scraping infrastructure. This removes a major technical barrier for businesses wanting to build AI applications that need real-time web data, potentially accelerating development of AI agents that can research, monitor competitors, or verify information autonomously.

We’re adding a web fetch tool to the Anthropic API 🪃 Using the web fetch tool, Claude will fetch and analyze content from any webpage URL—no additional infrastructure needed. https://x.com/alexalbert__/status/1965809009795153955

Stripe launches ChatGPT integration for direct payment processing
Stripe has released a Model Context Protocol (MCP) integration that allows ChatGPT users to process payments, check balances, and manage transactions directly within chat conversations. This marks a significant step toward AI assistants handling real financial transactions rather than just providing information, potentially transforming how businesses automate customer service and sales workflows. The integration demonstrates growing trust in AI systems to handle sensitive financial operations, though security and error-handling capabilities remain to be proven at scale.

Use @stripe in ChatGPT via our MCP”” / X https://x.com/emilygsands/status/1965824153996636423

Claude gains advanced Excel capabilities with formula generation
Anthropic’s Claude can now create and manipulate complex Excel spreadsheets, generating hundreds of formulas from single prompts and applying sophisticated formatting—a capability that surpasses other AI assistants in spreadsheet automation. This marks a significant advance in AI’s ability to handle structured data tasks that millions of knowledge workers perform daily, potentially transforming how businesses approach spreadsheet-based analysis and reporting.

Claude’s new ability to work with Excel files is the best I have seen so far I have given it existing spreadsheets to work with and asked it to create new ones. Good use of formatting, formulas, etc. It created all of this, including 406 formulas, from one prompt (& its solid). https://x.com/emollick/status/1965608685297922315

Claude creates complex Excel financial models from simple requests
A user with no financial modeling experience asked Claude to create a discounted cash flow model and received a complete 9-sheet Excel workbook with working formulas on the first try. This demonstrates AI’s ability to translate high-level requests into detailed, technical deliverables that previously required specialized expertise, potentially democratizing access to sophisticated financial analysis tools.

Claude please make a discounted cash flow financial model. Make no mistakes.”” I’ve never made a financial model in my life before this and Claude one shots a 9-sheet workbook complete with formulas for me. Vibe excel era. https://x.com/alexalbert__/status/1965828155744330022

Claude creates Excel art by replicating user profile pictures
Anthropic’s Claude AI demonstrated an unexpected creative capability by recreating a user’s profile picture entirely within Excel spreadsheet cells, suggesting AI assistants are expanding beyond traditional productivity tasks into artistic expression. This “vibe excel era” signals how AI is transforming mundane office tools into creative mediums, potentially changing how we think about both spreadsheet software and AI capabilities in workplace contexts.

Claude just replicated my profile pic in an excel file We’re entering the vibe excel era https://x.com/alexalbert__/status/1965792437504499803

OpenAI enables developers to build action tools directly in ChatGPT
OpenAI has launched Model Context Protocol (MCP) support in ChatGPT’s developer mode, allowing developers to create custom connectors that can perform write actions like updating Jira tickets or triggering workflows, not just retrieve information. This transforms ChatGPT from a conversational AI into a platform where developers can build and chain together automation tools, potentially replacing some dedicated workflow software for technical users.

We’ve (finally) added full support for MCP tools in ChatGPT. In developer mode, developers can create connectors and use them in chat for write actions (not just search/fetch). Update Jira tickets, trigger Zapier workflows, or combine connectors for complex automations. https://x.com/OpenAIDevs/status/1965807401745207708

AI agents gain ability to directly control user interfaces
A new open-source protocol called AG-UI enables AI agents to interact directly with frontend applications, completing the communication stack alongside existing agent-to-tool (MCP) and agent-to-agent (A2A) protocols. This matters because it removes a major bottleneck in AI deployment—agents previously couldn’t manipulate the visual interfaces humans use daily, limiting their practical applications in real-world workflows.

AI agents can finally talk to your frontend! The AG-UI Protocol bridges the critical gap between AI agents and frontend apps, making human-agent collaboration seamless. MCP: Agents to tools A2A: Agents to agents AG-UI: Agents to users 100% open-source. https://x.com/akshay_pachaar/status/1963945302991450272

Adobe launches AI agents to automate marketing and customer experiences
Adobe released six specialized AI agents that automate tasks like audience creation, customer journey design, and website optimization, with 70% of eligible customers already using the system. The agents work through Adobe’s Experience Platform, which connects enterprise data to enable contextual actions, while new tools allow businesses to customize agents and integrate third-party systems through partnerships with firms like Google Cloud and PwC.

Adobe Announces General Availability of AI Agents for Businesses to Transform Customer Experience Orchestration https://news.adobe.com/news/2025/09/adobe-announces-general-availability-ai-agents

Microsoft launches scripted audio mode for Copilot voice generation
Microsoft’s new MAI-Voice-1 model in Copilot Labs now offers three distinct audio generation modes: scripted (reads text exactly as written), emotive (adds dramatic interpretation), and story (performs multiple character voices). This gives users precise control over AI voice output, addressing a common complaint that AI voices often improvise or deviate from provided scripts when exact reproduction is needed.

You asked, we shipped! Scripted mode just dropped for audio generation in Copilot Labs (c/o our new MAI-Voice-1 model). Scripted mode: reads your input verbatim Emotive: riffs a bit for max drama Story: performs multiple voices/characters Try out all 3 ➡️ https://copilot.microsoft.com/labs/audio-expression

Chinese AI models crack top 10 on global leaderboard
Alibaba’s Qwen3-max-preview reached #6 and Kimi’s K2-0905-preview tied for #8 on the leading text AI benchmark, marking a shift as Chinese companies challenge Western dominance in large language models. The achievement signals China’s growing competitiveness in AI development despite export restrictions on advanced chips.

🚨 Leaderboard Disrupted! Two new models have entered the Top 10 Text leaderboard: 🔸#6 Qwen3-max-preview (Proprietary) by @Alibaba_Qwen 🔸#8 Kimi-K2-0905-preview (Modified MIT) by @Kimi_Moonshot tied with 7 others. Note that this puts Kimi-K2-0905-preview in a tight race for https://x.com/arena/status/1965115050273976703

Kimi’s open-source model breaks into elite coding benchmark territory
Chinese startup Kimi Moonshot’s K2-0905 model became the first open-source AI to score above 90% on Roo Code evaluations, ranking 7th overall at 94% accuracy while being the fastest and cheapest option in the top 10. This breakthrough challenges the dominance of closed proprietary models from major tech companies, potentially democratizing access to high-performance coding AI that previously required expensive commercial licenses.

📊 @Kimi_Moonshot’s K2-0905 on @GroqInc scored 7th overall at 94% on Roo Code evals, the 1st open-source model to break the 90+ barrier. It’s also the fastest and cheapest in the top 10, while holding its own on accuracy. View the full leaderboard: https://x.com/roo_code/status/1965098976677658630

Kimi K2 upgrade delivers major boost to AI agent capabilities
Chinese AI company Kimi’s latest K2 0905 model shows only modest gains in general intelligence (up 2 points on benchmarks) but demonstrates substantial improvements in agentic capabilities—the ability to autonomously complete multi-step tasks. This shift toward practical task completion over raw intelligence scores reflects the industry’s growing focus on making AI systems that can actually get work done rather than just score well on tests.

Kimi K2 0905 upgrade: Substantial improvement in agentic capabilities, modest change in overall intelligence Key takeaways: ➤ Intelligence increased +2 pts in our Artificial Analysis Intelligence Index ➤ Agentic capabilities substantially improved as shown by our two new https://x.com/ArtificialAnlys/status/1965010554499788841

Chinese AI models slash costs while matching Western performance
Chinese developers have released coding AI agents that match or exceed Claude Sonnet and GPT-4’s capabilities at dramatically lower prices—GLM 4.5 costs just $3 monthly while Kimi K2.1 runs three times faster and seven times cheaper than Anthropic’s Opus. This shift to high-quality, open-weight models from China could democratize access to advanced AI coding tools and disrupt the pricing models of major Western AI companies.

It feels the coding agent frontier is now open-weights: GLM 4.5 costs only $3/month and is on par with Sonnet Kimi K2.1 Turbo is 3x speed, 7x cheaper vs Opus 4.1, but as good Kimi K2.1 feels clean. The best model for me. GPT-5 is only good for complicated specs — too slow.”” / X https://x.com/Tim_Dettmers/status/1965021602267217972

Google brings Gemma 3n AI assistant to Android phones offline
Google has released Gemma 3n in the Play Store, enabling fully offline AI assistance with speech, text, and image capabilities on Android devices. This marks a significant shift toward privacy-focused, internet-free AI tools for consumers, with the model supporting real-time speech-to-text translation and processing audio clips up to 30 seconds—bringing ChatGPT-like functionality directly to phones without cloud dependency.

Gemma 3n now available in the Play Store for on-device, internet free with speech, text and image input! Open local AI Assistants are coming to everyone! 🤯 – New on-device speech-to-text and speech-to-translated-text. – Process audio batch inference for clips up to 30 seconds. https://x.com/_philschmid/status/1965742109157188031

Sony’s Hawk-Eye tracks stadium action in real-time 4D
Sony’s Hawk-Eye system uses dozens of stadium cameras to create a real-time spatial understanding of sports action, accurately tracking ball flight paths and player movements in four dimensions. This technology bridges physical and digital worlds for sports analytics, enabling instant replay decisions and performance insights that are transforming how professional sports are officiated and analyzed.

Connecting the world of bits and atoms has never been easier. This is Sony’s “”Hawk-Eye”” system that fuses dozens of cameras placed around a stadium to build a spatial 4D understanding of the action in real-time! It can accurately track the flight path of a ball, determining https://x.com/bilawalsidhu/status/1964053826878517556

OpenAI’s GPT-5 Pro shows expert-level medical diagnostic capabilities in trials
OpenAI is testing GPT-5 Pro with physicians who report its diagnostic abilities match those of top subspecialists at leading medical centers like Mayo Clinic. This marks a significant leap from current AI medical tools, potentially democratizing access to expert-level medical consultation, though details on accuracy rates and deployment timeline remain undisclosed.

gpt-5 pro as an aide to physicians — “as if the best sub specialist at the speciality centers like Mayo, had been given this case to look at””” / X https://x.com/gdb/status/1964401035419799983

Google releases NanoBanana image model to Gemini API free tier
Google has made its NanoBanana image generation model temporarily available for free through the Gemini API under the name “gemini-2.5-flash-image-preview” for weekend experimentation. This marks a shift in Google’s API strategy by offering advanced image generation capabilities without cost barriers, potentially accelerating developer adoption and creative applications ahead of the competitive holiday AI release season.

Good news: @NanoBanana is now available in the Gemini API free tier 🍌 for the weekend (under “gemini-2.5-flash-image-preview”). Go build something people want!”” / X https://x.com/OfficialLoganK/status/1964127721711157600

Google’s Nano Banana model automates mass ad creation at scale
A new n8n AI agent leverages Google’s Nano Banana image model to generate thousands of static advertising variations automatically, addressing the model’s single-image limitation. This marks a shift from one-off AI image generation to industrial-scale creative production, potentially transforming how businesses approach digital advertising campaigns by eliminating the manual work of creating multiple ad variants.

Google Nano Banana 🍌 is crazy good at static ads… But it only generates one image at a time. This n8n AI Agent helps you generate 1000s of winning ad variations in minutes, fully automated. → Built with the latest Nano Banana image model → Creates static ad images in https://x.com/mikefutia/status/1963967610611003671

ByteDance’s Seedream 4.0 tops AI image generation leaderboards
ByteDance’s new Seedream 4.0 model has overtaken Google’s Gemini 2.5 Flash to become the top-ranked AI for both text-to-image generation and image editing, according to Artificial Analysis benchmarks. The model shows particular strength in creating complex visual content like mathematical diagrams and infographics, demonstrating ByteDance’s growing competitiveness against Western tech giants in generative AI.

Bytedance’s answer to Nano Banana is really good at making high density infographics. “Draw the following system of binary linear equations and the corresponding solution steps on the blackboard: 5x + 2y = 26; 2x -y = 5.” “Create an infographic showing the causes of https://x.com/bilawalsidhu/status/1965838191019307476

Seedream 4.0 is the new leading image model across both the Artificial Analysis Text to Image and Image Editing Arena, surpassing Google’s Gemini 2.5 Flash (Nano-Banana), across both! Seedream 4.0 is the latest release from Bytedance Seed, and is a substantial improvement on https://x.com/ArtificialAnlys/status/1966167814512980210

OpenAI and Microsoft sign non-binding agreement amid partnership tensions
OpenAI and Microsoft announced a memorandum of understanding to continue their partnership, though reports suggest growing friction over the pace of AI development and compute resources. The non-binding nature of the agreement and concurrent reports of disagreements signal potential shifts in one of AI’s most significant corporate alliances, which has powered ChatGPT and shaped the industry’s competitive landscape.

A joint statement from OpenAI and Microsoft | OpenAI OpenAI and Microsoft have signed a non-binding memorandum of understanding (MOU) for the next phase of our partnership. https://openai.com/index/joint-statement-from-openai-and-microsoft/

OpenAI, Microsoft Rift Hinges on How Smart AI Can Get – WSJ https://www.msn.com/en-us/money/other/openai-microsoft-rift-hinges-on-how-smart-ai-can-get/ar-AA1Hqxoq

Oracle and OpenAI announce $300 billion cloud infrastructure partnership
Oracle has secured a massive $300 billion cloud computing deal with OpenAI to provide infrastructure for training and deploying AI models, marking one of the largest enterprise cloud contracts ever signed. The partnership signals OpenAI’s enormous compute needs as it scales ChatGPT and develops more advanced AI systems, while establishing Oracle as a major player in the AI infrastructure market alongside Microsoft Azure and Amazon Web Services. Oracle’s stock surged following the announcement, reflecting investor confidence that the company can capture significant revenue from the AI boom through its cloud services.

Exclusive | Oracle, OpenAI Sign $300 Billion Cloud Deal – WSJ https://www.wsj.com/business/openai-oracle-sign-300-billion-computing-deal-among-biggest-in-history-ff27c8fe?mod=hp_lead_pos1

Oracle Stock Skyrockets as Software Giant Scores Massive AI Deals – WSJ https://www.wsj.com/business/earnings/oracle-stock-orcl-ai-deals-047216cd

Tesla board proposes pay package that could make Musk a trillionaire
Tesla’s board unveiled a compensation plan granting CEO Elon Musk up to 423.7 million shares worth $143.5 billion today, but only if the company reaches an unprecedented $8.5 trillion market cap—nearly eight times its current $1.1 trillion value. The package aims to keep Musk focused on Tesla rather than his other ventures like xAI and SpaceX, as he currently owns just 13% of shares and has threatened to develop AI products elsewhere without 25% voting control. Critics argue the targets require delivering on long-promised but unmet goals like full self-driving capabilities, while the company faces growing competition from Chinese EV makers and loss of regulatory credit revenue.

Tesla proposes Elon Musk pay package that could make him the world’s first trillionaire | CNN Business https://edition.cnn.com/2025/09/05/business/elon-musk-tesla-new-pay-package

OpenAI’s Sam Altman credits two key technical leaders publicly
OpenAI CEO Sam Altman singled out Jakub Pachocki (Chief Scientist) and Szymon Sidor (technical leader) as essential to the company’s existence, a rare public acknowledgment of specific individuals beyond the executive team. This highlights how OpenAI’s success depends on key technical talent who typically work behind the scenes, underscoring the critical role of senior engineers in the AI race where companies compete fiercely for top researchers.

There are two people I’d like to mention that OpenAI would not be OpenAI without: Jakub Pachocki and Szymon Sidor. https://x.com/sama/status/1965206375287672834

AI search startup Perplexity hits $20 billion valuation
Perplexity raised $200 million at a $20 billion valuation, just two months after securing funding at $18 billion, as the AI-powered search engine approaches $200 million in annual recurring revenue. The rapid growth and aggressive fundraising pace underscore intense competition in AI search, with Perplexity positioning itself as a Google challenger—even offering to buy Chrome for $34.5 billion amid antitrust proceedings.

Perplexity reportedly raised $200M at $20B valuation | TechCrunch https://techcrunch.com/2025/09/10/perplexity-reportedly-raised-200m-at-20b-valuation/

Chinese robotics maker Unitree seeks $7 billion IPO valuation
Unitree, known for its quadruped robots competing with Boston Dynamics, is preparing one of China’s largest tech IPOs in years with a Q4 listing on Shanghai’s STAR Market. The company’s valuation would jump over 4x from its last funding round at $1.7 billion, signaling strong investor appetite for robotics firms despite the broader tech downturn.

Unitree is targeting a $7B valuation for its IPO, planned for Q4 this year. The listing, likely on Shanghai’s STAR Market, would be one of the biggest onshore tech listings in recent years It was valued at CN¥12B ($1.7B) in the last funding round. https://x.com/TheHumanoidHub/status/1965273782899355842

AI vendors cut corners on model quality to boost profits
A growing number of software-as-a-service companies are choosing cheaper, less capable AI models and outdated implementation techniques to maximize margins rather than deliver the best possible AI features to customers. This creates a conflict where vendors prioritize cost savings over performance improvements, even as better AI capabilities become available, potentially leaving enterprise customers with subpar AI tools that lag behind the state of the art.

The problem with using SaaS vendors with their own AI solutions is that their incentives are to use cheap models, as little reasoning as possible & to stick with outdated prompting & RAG strategies than updating them as AI improves Not all vendors succumb to temptation, many do.”” / X https://x.com/emollick/status/1965204136984805399

Perplexity launches free AI search for US government workers
AI search startup Perplexity released a government-specific version offering secure access to premium AI models without requiring contracts, logins, or enterprise licenses. The move targets federal employees who need advanced research capabilities but face procurement barriers and security requirements, potentially accelerating AI adoption across government agencies that have lagged behind private sector usage.

Introducing Perplexity for Government: zero data usage, fully secure, access to premium AI models, deep research, no logins, no enterprise contracts or licenses. 🇺🇸🇺🇸🇺🇸”” / X https://x.com/AravSrinivas/status/1965032305053065590

Today we are rolling out Perplexity for Government. We believe that America’s government deserves the best AI, no contract required. Starting today, all U.S. Government access to Perplexity is now secure by default with advanced model selection included. https://x.com/perplexity_ai/status/1965030156415980009

Microsoft AI head warns against building seemingly conscious AI
Microsoft AI’s leader argues that AI systems mimicking consciousness pose immediate societal risks, even without being truly conscious. He predicts that within 2-3 years, existing technology could create “Seemingly Conscious AI” that appears so lifelike people will demand rights and protections for it, creating dangerous delusions and social divisions. His solution: build AI explicitly as tools to empower humans, not as simulated beings, establishing clear norms before the technology arrives.

Could AI ever be conscious?”” is a distraction from the here and now. We’re on the brink of Seemingly Conscious AI and that illusion will only escalate delusions and dependence. We need to build AI to better human lives, not to cosplay living its own. https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming

There’s Only One Lonely AI Visual: Week Ending September 12, 2025

Ominous rocks. One of the more interesting levers to use in creating AI images is Midjourney’s style codes, which let you take the style of an image and share it. These are all the same prompt, all done as photographs (rather than drawings), but with different style codes. https://x.com/emollick/status/1965835264229523788