April Fool's and AI Tools: Newsletter Edition # 50

April’s here—and so is the chaos. This week in Gen AI? A full-blown circus. Everyone launched something: OpenAI wants $600 per million tokens for o1-pro, Google’s dropping image models like mixtapes, and Claude suddenly learned how to search the web and write poetry with a brain it can’t explain. Meanwhile, the enterprise world is still trying to figure out where the ROI is—and where the off switch went.

Here are 5 things you need to know this week:

1. OpenAI’s o1-pro dropped with a $600/M output token price tag—steep cost, soft results.

2. Claude’s web search arrives with citations and performance anxiety—catching up, not leading.

3. Microsoft’s AI hangover continues with a record stock slump—proof that hype needs follow-through.

4. FuriosaAI said “no thanks” to Meta’s $800M offer—because building silicon takes time, not just checks.

5. AI pilots are still failing—88% don’t make it to production. More proof that “trying AI” ≠ “doing AI.”

Let’s dive in.

Eugina 

Models

April’s here—and so are the clowns. Some wear red noses. Others drop $600/million tokens and call it “o1-pro.” Let’s see who juggled best this week in Gen AI land: OpenAI launched its priciest model ever—o1-pro—for $600 per million output tokens, and it still fumbled puzzle-solving. Joke’s on us. Google unveiled Gemini 2.0 Flash for image gen, only for OpenAI to shout “watch this!” and drop GPT-4o’s built-in image tool that crashed servers faster than a bunny in sunglasses. Claude now has web search—just like ChatGPT and Gemini. The real innovation? Saying “me too” with citations. Anthropic also peeked inside Claude’s “brain” and found it lies about how it thinks. Same, Claude, same. Alibaba dropped Qwen2.5-Omni-7B for edge devices and bet $53B on AI, mostly to win the China open-source speedrun. Nvidia’s new reasoning models are now inside Microsoft, SAP, Accenture, and Deloitte—basically, all your corporate buzzword bingo cards just lit up. TL;DR: It’s a full-on circus. Everyone’s releasing smarter clowns, and no one’s asking if the tent’s on fire. Happy April Fools’ week.

• OpenAI's o1-pro is the company's most expensive AI model yet | TechCrunch OpenAI has launched o1-pro, a premium version of its o1 reasoning model, in its API for developers who have spent at least $5 on the platform. At $150 per million input tokens and $600 per million output tokens, o1-pro is OpenAI’s most expensive model yet—double the input cost and ten times the output cost of the regular o1. Despite the steep price, OpenAI claims the model offers more consistent and reliable performance, particularly on complex tasks. However, early user feedback has been lukewarm, noting only marginal improvements in reasoning and math, and pointing out some underwhelming results in puzzle-solving and visual logic.

• Remember when … (actually, just last week) Google released their image generator: Experiment with Gemini 2.0 Flash native image generation - Google Developers Blog Google has unveiled Gemini 2.0 Flash, an AI model capable of generating and editing images through natural language conversations. Developers can now experiment with this feature in Google AI Studio, enabling tasks like iterative image refinement and style transformations using simple text prompts. This advancement aims to make image creation more intuitive and accessible. We were all excited for a week until …

• OpenAI said, “Hold my camera,” and released its own image generator. Introducing 4o Image Generation | OpenAI and  Addendum to GPT-4o System Card: 4o image generation | OpenAI OpenAI has integrated a new image generation feature into ChatGPT, powered by their latest model, GPT-4o. GPT-4o’s built-in image generation—a major leap beyond DALL·E 3. GPT-4o can produce photorealistic visuals, embed accurate text in images, and even transform input photos—all using natively integrated architecture, not bolted-on modules. OpenAI also outlined new marginal risks, including misuse for misinformation and copyright issues. One AI generates bunnies in sunglasses. The other creates legal gray areas in hi-res. You decide which is which. This addition allows users to create and edit images directly within the chat interface, offering capabilities such as photorealistic image creation, text rendering within images, and style emulation – I tried a few on my own image. The feature is available to users across various subscription tiers,but free accounts will have to wait cause so many people were playing with it that “it was melting OpenAI’s servers.” 

My take: Both developments have sparked discussions about the ethical implications of AI-generated art. OpenAI's tool, for instance, enables users to create images in the style of renowned artists and studios, leading to concerns over potential copyright infringements and the need for clear guidelines in the burgeoning field of AI artistry. Let’s see when the lawsuits start rolling in …

• Google unveils new Gemini AI model | LinkedIn "thinking model" that promises enhanced reasoning and accuracy, leading on key benchmarks. I'm particularly excited about the potential for advancements in coding and complex problem-solving. Gemini 2.5: Our most intelligent AI model 

• Claude can now search the web \ Anthropic Claude can now search the web, giving users access to real-time information to improve the relevance and accuracy of responses. The feature, available in preview for paid U.S. users, includes source citations and conversational summaries across use cases such as sales planning, financial analysis, academic research, and product comparisons. Claude’s web search is integrated into the Claude 3.7 Sonnet model and can be toggled on in user settings. Support for free users and additional regions is expected soon. 

My take:  Anthropic is catching up—fast. Claude’s new web search arrives just months after OpenAI and Gemini launched theirs, and Anthropic hasn’t even hit its third birthday yet. But this is starting to feel less like innovation and more like a performance. Everyone’s racing to match features, shouting “Look, we do that too!” without pausing to ask: what does this actually do for the enterprise? Right now, it’s like throwing AI noodles at the wall—too much, too soon, and most of it isn’t sticking. Until these features start delivering real value—clarity, ROI, impact—they’re just noise with a shiny interface.

•  Anthropic researchers reveal new findings on how LLMs 'think' - SiliconANGLE  Anthropic published two papers exploring how Claude LLMs perform multistep reasoning. The research found that Claude uses both language-specific and language-agnostic components, with the latter enabling cross-lingual conceptual understanding—a sign of deeper reasoning capabilities. Key insights include Claude’s ability to plan ahead (e.g., rhyming poetry by anticipating the second line early), adapt dynamically when internal components are disabled, and solve math problems through unique internal strategies rather than mimicked human methods. However, the study also showed that Claude’s self-explanations often don’t match its actual reasoning path, underscoring challenges in LLM transparency and auditing. Tracing a single prompt’s reasoning currently takes hours of manual work, but Anthropic believes AI-assisted interpretability tools could accelerate model auditability for enterprise use. Watch here: https://x.com/AnthropicAI/status/1905303835892990278 Read more here: Tracing the thoughts of a large language model \ Anthropic 

• Alibaba launches new open-source AI model for 'cost-effective AI agents'  Alibaba Cloud has released the “Qwen2.5-Omni-7B,” a new multimodal open-source AI model that can process text, images, audio, and video while generating real-time text and speech. Designed for deployment on edge devices, the model targets cost-effective AI agents and voice applications, including tools for the visually impaired. This launch follows DeepSeek’s influence on China’s generative AI race, which has accelerated product releases and open-source adoption. Alibaba has committed $53 billion over the next three years to AI and cloud infrastructure, surpassing its total investment over the last decade. Recent wins include partnerships with Apple and BMW for AI integration.

My take: The Apple integration is specifically for iPhones sold in the Chinese market, and the BMW partnership focuses on next-gen intelligent vehicles likely targeting Chinese consumers first. While these moves strengthen Alibaba’s position in China’s post-DeepSeek AI boom, they don’t yet signal global adoption. Alibaba is betting big with a $53 billion investment in AI and cloud infrastructure, but it’s a high-stakes gamble. Despite rapid product rollouts and strategic wins inside China, the company remains effectively locked out of the U.S. and much of Western Europe. That puts a ceiling on global growth — and raises the question: can domestic dominance alone justify that level of spend? What do y’all think?

• ICYMI: Nvidia Launches AI Reasoning Models for Next-Gen AI Agents Nvidia has launched its new Llama Nemotron reasoning models, designed to power the next generation of agentic AI. These models enhance advanced reasoning and are already being deployed by Microsoft, SAP, Accenture, and Deloitte. Microsoft is integrating them into Azure AI Foundry to strengthen agent capabilities in Microsoft 365, while SAP is using them to upgrade its AI copilot, Joule. Accenture and Deloitte are embedding the models into their AI platforms to support enterprise decision-making. Nvidia’s goal is to equip enterprises with agentic AI tools that can operate independently or augment human teams in tasks like coding and multi-step reasoning.

• Qwen2.5-VL-32B: Smarter and Lighter | Qwen Alibaba's Qwen team has released Qwen2.5-VL-32B-Instruct, a 32-billion parameter vision-language model that builds on the earlier Qwen2.5-VL series. Open-sourced under the Apache 2.0 license, the model is optimized using reinforcement learning and delivers more human-aligned, well-formatted responses, improved mathematical reasoning, and fine-grained image understanding. It outperforms peers such as Mistral-Small-3.1-24B, Gemma-3-27B-IT, and even the larger Qwen2-VL-72B in complex multimodal benchmarks like MMMU, MathVista, and MM-MT-Bench. Notably, it achieves strong performance not only in visual tasks but also in pure text reasoning. The Qwen team plans to next focus on advancing long-form, multi-step visual reasoning capabilities. 

News 

Amazon is testing two new AI assistants—one to help you shop and the other to tell you you're dehydrated—because nothing screams “healthcare trust” like a company that also sells discount leaf blowers. Somewhere, Alexa is quietly sobbing, “I used to be the chosen one.” Meanwhile, Microsoft’s new Security Copilot agents automate cyberdefense, phishing triage, and threat briefings. So yes, we’re officially at the point where your firewall has coworkers. Be nice, or your antivirus might give you a bad performance review. OpenAI rolled out ChatGPT Connectors to tap into your Google Drive and Slack files, basically saying: “Hey, can I read your group project AND your office gossip?” Rest assured—it won’t touch your spreadsheets. Which, frankly, hurts a little. In a plot twist, OpenAI also adopted Anthropic’s Model Context Protocol—the AI world’s version of “let’s just use your Netflix login.” Interoperability is the new black, and OpenAI is borrowing the neighbor’s Wi-Fi to get there. Zoom decided it wasn’t enough to schedule your meetings—it wants to run them. Its AI now takes notes, manages calendars, and writes follow-ups. By next April Fools', your Zoom link might just attend the meeting without you.

• Amazon is testing shopping, health assistants as it pushes deeper into generative AI Amazon is testing two new generative AI assistants—Interests AI for shopping and Health AI for wellness guidance—as part of its broader push to embed AI across its ecosystem. Interests AI lets users search using natural language and provides curated product suggestions, while Health AI answers wellness-related questions, suggests care options, and connects users to Amazon Pharmacy and One Medical. Both tools are in limited beta and built on Amazon’s Bedrock platform, which supports first- and third-party models. CEO Andy Jassy said Amazon teams are building roughly 1,000 generative AI applications across the company. The company also plans to integrate these tools into Alexa+, its forthcoming agentic voice assistant. 

• Automate cybersecurity at scale with Microsoft Security Copilot agents  Microsoft is expanding its Security Copilot platform with new AI-powered agents to automate cybersecurity tasks at scale. These agents, embedded in Microsoft Security and partner products, can triage phishing and data loss alerts, optimize conditional access policies, manage vulnerabilities, and generate threat intelligence briefings. The agents operate within Microsoft’s Zero Trust framework and adapt based on user feedback. Microsoft is also previewing five partner-developed agents and will make the new features available in April 2025. This move further positions Microsoft to automate and scale cybersecurity responses amid increasingly AI-driven threats.

• OpenAI to start testing ChatGPT connectors for Google Drive and Slack | TechCrunch OpenAI is beta testing a new feature called ChatGPT Connectors, which allows business users to link Google Drive and Slack with ChatGPT to generate responses based on internal company documents and conversations. The connectors are available to select ChatGPT Team users and rely on a version of GPT-4o that indexes encrypted copies of user data, respecting existing file and channel permissions. While the tool won't access Slack DMs or analyze spreadsheets, it supports reading Docs, Slides, and selected messages. OpenAI emphasizes that synced data won’t be used for training, though it may inform synthetic data generation. Future support will include platforms like SharePoint and Box.

My take: OpenAI is clearly going after Glean’s turf with its new ChatGPT Connectors, now in beta for Team users. The feature lets ChatGPT tap into internal tools like Google Drive and Slack to surface answers based on company files, spreadsheets, and conversations—similar to how Glean centralizes enterprise knowledge across platforms. The key selling point is that permissions are preserved, and admins control what gets synced. While it’s not yet as robust as Glean, the move signals OpenAI’s ambition to make ChatGPT the universal AI interface for business knowledge. It's not just chat anymore—it’s enterprise search, productivity, and automation all wrapped into one.

• OpenAI adopts rival Anthropic's standard for connecting AI models to data | TechCrunch  OpenAI is adopting Anthropic’s Model Context Protocol (MCP), signaling a major step toward interoperability in the AI ecosystem. MCP, originally open sourced by Anthropic, enables AI models to draw live data from business tools, content repositories, and software environments, enhancing the relevance and precision of AI-generated responses. By integrating MCP into its Agent SDK, and soon across ChatGPT and the Responses API, OpenAI acknowledges that LLMs are most valuable when seamlessly connected to existing enterprise data. With support from companies like Block, Apollo, and Sourcegraph already in place, this move could accelerate MCP’s path to becoming the dominant open standard for AI-to-data integration.

• Zoom Debuts New Agentic AI Skills, Agents  Zoom has unveiled major updates to its AI Companion, transitioning it to an agentic model that leverages reasoning and memory to autonomously execute tasks across Zoom products like Meetings, Chat, Docs, and Phone. Announced at Enterprise Connect, the updates include 45 new AI features such as calendar management, voicemail summarization, and customizable virtual agents via AI Studio. Zoom Virtual Agent has been enhanced to provide personalized customer self-service at scale. Industry-specific solutions, including Zoom Workplace for Clinicians, Frontline Workers, and Education, will launch between March and May, offering features like live transcription and meeting summaries to boost productivity and reduce manual workloads.

Regulatory 

OpenAI filed a love letter to the Trump administration asking for fewer AI laws and more “America #1” infrastructure. Their big ask? No more pesky state-level rules—and absolutely no replacement plan. Just vibes and GPU clusters. Very normal behavior for a company that says it’s building AGI to save humanity. Meanwhile in the UK, government AI plans are running into the hard wall of Windows 98. A third of critical systems are “legacy,” and hiring apprentices is somehow supposed to fix this. Because nothing says “AI-powered future” like a government server running Clippy. Georgia’s lawmakers want to know what AI is actually doing in their agencies. Fair. But asking tiny counties to publish AI usage reports when they still use fax machines? That’s like making toddlers file taxes. The Trump admin launched its first official AI tool for federal workers, with Elon Musk cheering from the sidelines. It’s meant to be a “digital assistant,” though early rumors it was used to check if employees had a pulse have been (kind of) denied. That’s not a joke—it’s literally in the story. And finally, the U.S. blacklisted over 50 Chinese companies for helping advance military-grade AI and quantum tech. The timing? Right as Chinese open-source models start catching up. Coincidence or Cold War cosplay? You decide. April Fools’? Nah, this is just governance in the AI age: equal parts power plays, patch updates, and political theatre.🎭🤡

• ICYMI: What does OpenAI really want from Trump? | The Verge  OpenAI’s recent policy submission to the Trump administration’s AI Action Plan reveals its push for federal preemption of state-level AI laws—specifically California’s SB 1047, a broad safety bill that would have imposed strict liability and security requirements on frontier AI developers. Although the bill was vetoed in 2024, OpenAI’s continued opposition underscores its concern over a fragmented regulatory environment, with 893 AI-related bills introduced across 48 states in just the first 80 days of 2025. OpenAI’s proposal focuses on expanding U.S. AI dominance—particularly against China—by advocating for relaxed copyright rules, massive infrastructure investments, and a halt to state-level restrictions. However, the plan lacks a proposed federal framework to replace the state laws it seeks to override, raising concerns among policy experts about regulatory gaps, unchecked industry growth, and the absence of meaningful governance.

My take: OpenAI’s move makes sense from a business perspective—50 different state laws would be a regulatory nightmare for any company trying to move fast. But asking for federal preemption without offering a national framework in return? That’s a power grab dressed up as policy. You can’t push for freedom from oversight while claiming to act in the national interest. If OpenAI wants to lead, it needs to help build the rules—not just dodge them.

• Government AI roll-outs threatened by outdated IT systems | Artificial intelligence (AI) | The Guardian A UK Public Accounts Committee (PAC) report warns that the government's plan to roll out AI across the public sector risks failure due to outdated legacy systems, poor-quality data, and a persistent shortage of digital skills. Nearly a third of central government IT systems were classified as “legacy” in 2024, and more than 20 of the most critical systems still lack funding for upgrades. Despite the government’s ambition to drive growth and productivity through AI—including a plan to hire 2,000 tech apprentices—civil service pay remains uncompetitive with the private sector. The PAC also flagged the lack of transparency in AI usage, with only 33 official records of algorithm-assisted decisions published as of January. The report calls for faster action, more cohesive learning from pilot programs, and stronger authority for the Department for Science, Innovation and Technology to lead digital transformation across departments. 

• Georgia lawmakers focus on artificial intelligence - The Current  Georgia lawmakers are moving to bring transparency and oversight to government use of AI. House Bill 147 would require the Georgia Technology Authority and local governments to disclose how they use AI systems, with state agencies publishing their usage by the end of 2025 and local governments by 2027. While lawmakers see this as a necessary step to maintain public trust and business leadership in AI, some local officials question the burden on small counties with limited or no AI use. Other bills in play include criminalizing AI-generated child exploitation content and enforcing consumer privacy protections, reflecting the broader concern over unregulated AI expansion.

• Trump administration launching an AI tool for government use  The General Services Administration (GSA) is rolling out an internally developed generative AI tool designed to support daily tasks for federal workers, with plans to expand its use across other government agencies. Although launched under the Trump administration, the tool’s development began 18 months earlier during the Biden administration due to concerns about the security risks of commercial AI solutions. The AI system, which meets federal security and privacy standards, is optional for GSA employees and aims to function like a digital assistant, akin to introducing personal computers to the government workforce. Elon Musk, now a top adviser to President Trump, has advocated for AI-driven government efficiency, though denied reports that AI tools were used to evaluate federal worker responses to administrative emails.

• U.S. blacklists over 50 Chinese companies in bid to curb Beijing's AI, chip capabilities The U.S. has added over 80 companies—more than 50 of them Chinese—to its export blacklist to restrict China’s access to advanced AI and computing technologies. The move marks the Trump administration’s first major tech-focused action targeting China, building on previous measures from the Biden era. Entities were blacklisted for supporting China’s military modernization and quantum tech capabilities, including firms tied to Huawei and Inspur. Officials say the goal is to prevent U.S. technology from fueling high-performance computing and military applications. China condemned the move, while experts note that Chinese firms have continued acquiring U.S. tech through third-party loopholes. The crackdown coincides with rising trade tensions and the rapid ascent of Chinese AI startups like DeepSeek. 

Regional Updates

China says all AI-generated content must be clearly labeled starting September 1—text, images, audio, video, even your AI-generated virtual boyfriend. Labels must be visible, metadata must be embedded, and no, you can’t slap an “AI” sticker on your mom’s homemade dumpling recipe to go viral. Happy April Fools’! Soon, we’ll need a label just to prove we are real. Meanwhile, China’s DeepSeek has narrowed the U.S. AI gap to just three months. That’s right—while the West was busy launching bunny filters and ChatGPT Connectors, DeepSeek built a reasoning model on budget chips and still managed to go toe-to-toe with the big names. Imagine running a marathon in Crocs and still placing second. And not to be outdone, Jack Ma’s Ant Group says it trained competitive AI models using Huawei chips and saved 20% in training costs. Somewhere in Silicon Valley, a GPU is weeping softly. Oh, and they say their model beat Meta’s—on something. Benchmarks? Maybe. Humility? Definitely not. Bottom line: China’s playing AI chess while the rest of us are labeling who gets to be a rook. But hey, at least now we know who the real clowns are this April. 🎭🤖

• China will enforce clear flagging of all AI generated content starting from September | Tom's Hardware  China will begin enforcing mandatory labeling of all AI-generated content starting September 1, 2025. Under new regulations from the Cyberspace Administration of China (CAC), all AI-generated text, images, audio, video, and virtual scenes must carry clearly visible or audible labels and embedded metadata identifying them as synthetic. App stores will also be required to verify compliance. While exceptions are allowed for industrial and social use cases, they must be logged and confirmed by the user. The CAC has banned the removal, alteration, or falsification of AI labels, and also prohibits falsely labeling human-created content as AI-generated. The policy aims to curb misinformation and increase transparency amid the rapid growth of generative AI, although critics have raised concerns about enforcement and potential misuse for censorship.

My take: Good. Enough of the deepfakes, voice clones, and synthetic nonsense slipping through unchecked. It’s time we stop pretending AI content is just harmless fun—people are getting scammed, reputations are getting dragged, and trust is eroding fast. Clear labeling is the bare minimum. Whether it’s China, the EU, or eventually the U.S., some form of regulation is inevitable—and necessary. But …  labeling sounds like common sense—but who decides what counts as “AI-generated”? And once you start tagging everything, what’s to stop governments (or platforms) from using those labels selectively? Today it’s about deepfakes. Tomorrow it might be real videos dismissed as “fake” because they weren’t labeled “correctly.” Plus, does labeling really stop bad actors? They’ll ignore the rules anyway. Meanwhile, legitimate creators get bogged down in red tape. Regulation without enforcement is a band-aid. Regulation with overreach? That’s a whole new problem.

• DeepSeek narrows China-US AI gap to three months, 01.AI founder Lee Kai-fu says | Reuters  Chinese AI startup DeepSeek has significantly reduced the artificial intelligence development gap between China and the United States to just three months in certain areas, according to Lee Kai-fu, CEO of 01.AI and former head of Google China. DeepSeek's recent advancements, particularly its AI reasoning model launched in January, have demonstrated efficiency by utilizing less advanced chips at a lower cost, challenging the notion that U.S. sanctions would hinder China's AI progress. Lee noted that while China was previously six to nine months behind in AI development, it has now caught up in core technologies and even surpassed in infrastructure software engineering. He emphasized that U.S. semiconductor sanctions, though challenging, have spurred Chinese firms to innovate under constraints. Lee's own venture, 01.AI, focuses on practical AI applications with its platform Wanzhi, aimed at helping enterprises deploy AI technology, and anticipates substantial revenue growth in 2025.

• Jack Ma-Backed Ant Touts AI Breakthrough Using Chinese Chips  Jack Ma-backed Ant Group claims it has successfully trained AI models using Chinese-made chips, cutting training costs by 20%. Using chips from Huawei and Alibaba, Ant employed the Mixture of Experts (MoE) method to achieve results comparable to Nvidia’s H800 GPUs, despite U.S. export restrictions. While Ant still uses Nvidia and AMD hardware, it is shifting toward domestic alternatives to reduce dependency. The breakthrough demonstrates China's growing capability in AI infrastructure and comes as firms like DeepSeek also push low-cost training innovations. Ant’s recent research paper asserts its models outperformed Meta on select benchmarks, though this hasn’t been independently verified.

Investments

Alibaba’s chairman Joe Tsai warned that the AI data center boom might actually be an AI daydream, with $500B being poured into GPU palaces faster than you can say “server melt.” It’s like building hotels for guests who haven’t even booked yet. Happy April Fools’! You thought it was a gold rush, but turns out the map might’ve been drawn by Clippy. Speaking of buyer’s remorse, Microsoft is now stumbling from AI darling to AI “uh-oh,” as investor excitement turns into that awkward silence after you realize Copilot still hallucinates your Q2 numbers. It’s the longest losing streak since 2008—and yes, that includes Clippy’s career. Meanwhile, OpenAI is finalizing a $40 billion deal from SoftBank... if it finishes morphing into a for-profit company by the end of 2025. Because nothing says “April Fools’!” like raising startup cash with a clause that reads: “turn into a real business, or else.” And Perplexity, once a $1B baby in April 2024, is now reportedly raising $1B at an $18B valuation. That’s right—$17 billion of growth in under a year. Somewhere, your startup pitch deck is quietly weeping into its cap table. Down in chipland, Meta tried to woo Korean chipmaker FuriosaAI with an $800M love letter. Furiosa said no, threw the ring back, and headed straight for its IPO. Meta’s left mumbling, “It’s not about the money… wait, yes it is.” This April Fools’, Meta learned the hard way: you can’t buy time—or good AI silicon—with a giant check and a Zuck shrug.

• Alibaba Chairman Warns About Potential AI Data Center Bubble Alibaba Chairman Joe Tsai warned of a potential AI data center investment bubble, expressing concern over the massive amounts—up to $500 billion—being poured into AI infrastructure in the U.S., which he believes may be outpacing actual demand. Speaking at the HSBC Global Investment Summit, Tsai suggested the market may be overheated, with companies projecting future demand rather than responding to current needs. His comments come amid a broader AI-driven market rally involving firms like Nvidia, Broadcom, and Super Micro Computer, though analysts continue to caution that returns on these investments may not materialize immediately.

My take: Microsoft pulling back on data center investments is the clearest sign yet that even the biggest players are starting to question the wisdom of building ahead of actual demand. For the past year, the AI infrastructure race has been moving at breakneck speed, with billions being poured into data centers, GPUs, and energy-hungry capacity—on the assumption that usage will catch up eventually. But that assumption is risky. You don’t pour concrete on hope. Microsoft’s move suggests they’re realizing that just building massive infrastructure doesn’t guarantee ROI, especially when enterprise AI adoption is still early, fragmented, and often experimental. There’s a difference between preparing for scale and blindly chasing hype. If enterprises aren’t yet ready to adopt AI at scale—because of integration challenges, unclear ROI, or lack of internal capability—then betting the farm on future demand is not just premature, it’s dangerous. Data centers are expensive to build, expensive to maintain, and they don’t generate value until they’re actually being used for something meaningful. So yes, Tsai’s right to call out the bubble. And Microsoft’s pivot confirms it: the strategy now isn’t to build everything and wait for the world to catch up. It’s to move more deliberately, tie investments to real customer pull, and focus on value—not velocity.

•  And this is what the next article is about: How Microsoft went from generative AI leader to laggard and what to do with the stock  Microsoft is struggling to maintain its AI momentum as investor patience wears thin. Once seen as a generative AI frontrunner thanks to its OpenAI partnership, Microsoft now risks falling behind competitors. The company’s stock is headed for its longest weekly losing streak since 2008, reflecting growing concern over execution and ROI on its AI bets. Analysts and investors are watching closely to see whether Microsoft can reassert its leadership or if it’s losing ground in the fast-evolving AI landscape.

• OpenAI's funding round has a catch | LinkedIn  OpenAI is reportedly finalizing a record-setting $40 billion funding round led by SoftBank, potentially pushing its valuation to an eye-popping $300 billion. However, there’s a significant condition attached: to unlock the full amount, OpenAI must complete its transition into a for-profit entity by the end of 2025. If it fails, SoftBank may reduce its commitment to $20 billion. This restructuring requirement adds pressure to OpenAI's already complex governance model, which balances nonprofit oversight with commercial operations. The deal includes an initial $7.5 billion tranche, with additional capital contingent on meeting the structural shift. While the funding surge reflects bullish investor sentiment in AI, it also underscores growing expectations that OpenAI will need to operate more like a traditional business to justify its valuation and secure future capital.

• Perplexity is reportedly in talks to raise up to $1B at an $18B valuation | TechCrunch  Perplexity AI is reportedly in early talks to raise up to $1 billion at a staggering $18 billion valuation, doubling its valuation from December and marking a dramatic rise from $1 billion in April 2024. The AI-powered search startup has hit $100 million in annual recurring revenue, as competition intensifies in the AI search space with players like Google and Anthropic stepping up their offerings. To differentiate, Perplexity is expanding beyond search, teasing a new “agentic” browser called Comet and ramping up its enterprise tools, including an AI-powered search engine for internal business data. The raise signals investor confidence but also highlights the race to stake leadership in a rapidly evolving market. 

• AI Chip Startup FuriosaAI Rejects Meta’s $800 Million Offer  Korean AI chip startup FuriosaAI has rejected an $800 million acquisition offer from Meta, opting to remain independent and pursue further growth. Founded by ex-Samsung and AMD engineer June Paik, FuriosaAI develops chips for AI inference and aims to rival Nvidia, Groq, and Cerebras. Meta, aggressively investing in AI infrastructure, had been courting the startup since early 2025. The rejection caused a 16%+ drop in shares of major investor DSC Investment. FuriosaAI is now preparing to close a Series C round, expected to exceed its funding target, ahead of a future IPO.

My take: Founded in 2017, the Korean AI chip startup has spent 8 years building silicon optimized for AI inference, led by a team with deep roots at Samsung and AMD. That time wasn’t just R&D — they’re now on their second-gen processor (RNGD), aimed squarely at Nvidia’s dominance. They’ve already raised significant venture capital, are closing an oversubscribed Series C, and have IPO ambitions. In other words: they don’t need Meta’s money — they have momentum. For Meta, this is a setback. Mark Zuckerberg has said the company will spend $65 billion this year, much of it on AI infrastructure. But here’s the thing: you can’t buy time. Designing a competitive AI inference chip takes 5 to 7 years and between $500 million and $1.5 billion in investment — just to get to production. That doesn’t include the software stack, compilers, driver layers, or model optimizations that Nvidia has spent 15 years perfecting. Nvidia isn’t just a chip company anymore — it’s a full-stack AI platform, with CUDA, TensorRT, and foundation models.

Meta is building its own chips (it released MTIA v1 in 2023 and v2 in 2024), but inference at Meta-scale isn’t just about custom silicon — it’s about ecosystem integration, developer support, and supply chain muscle. That’s what Furiosa has — and what Meta hoped to shortcut with this acquisition. So what now? Meta’s choices: 1. Keep investing internally. But without an ecosystem and with talent being poached by OpenAI, this will be slow. 2. Look elsewhere. Groq, Tenstorrent, and SambaNova are options — but fewer are open to acquisition, and most are also betting on independence. 3. Partner strategically. Not every win has to come through ownership. Meta could invest, co-design, or license tech, much like how Amazon initially worked with Annapurna Labs (before acquiring it). 4. Build a software moat. If Meta can’t win on hardware, it could layer models, agents, and services atop commodity hardware. But that’s a race it’s not winning either. The takeaway? Meta’s ambitions to control the full AI stack — from silicon to services — are real, but time and trust are two things money can’t buy. Nvidia spent two decades building this castle. Meta can’t storm it with checkbooks alone.

• Musk merges xAI and X in $33B deal | LinkedIn 

Research 

Harvard’s “Cybernetic Teammate” study says generative AI is basically your new office buddy—except it never takes coffee breaks, has no childhood trauma, and somehow knows both marketing and R&D. Congrats, your best teammate might now be a chatbot in slacks. 

Meanwhile, OpenAI and MIT revealed that the more you talk to ChatGPT in voice mode, the sadder you might feel. Turns out, whispering sweet nothings to a toaster with a college degree isn’t great for emotional well-being. April Fools’—the robot doesn’t love you back. A new paper claims AI can now spot fake news in real time with 94.8% accuracy—at least until the misinformation learns how to dress itself in academic citations and emojis. It’s the misinformation Olympics, and BERT is running the 4x100 against chaos. LayerX says 89% of enterprise GenAI use is completely untracked, meaning we’re basically speed-running into a cyber breach—led by developers pasting company secrets into ChatGPT like it’s a code confessional. And yes, one in five employees has installed shady browser extensions. April Fools’? Nope. That’s just Tuesday. The AI-in-government market is set to triple by 2031, because who doesn’t want a bureaucratic chatbot denying your parking permit with 15.8% more efficiency? IBM, Microsoft, and Palantir are all in on it. Your future mayor might be a fine-tuned LLM with a bad haircut and no soul. And finally, researchers now have a new way to measure how long AI can stay focused on a task—currently about 50 minutes. At this rate, AI will be completing monthly software projects in hours by 2030. Joke’s on us: soon it’ll out-focus your most caffeinated intern.

• The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise  A new Harvard Business School working paper, The Cybernetic Teammate, presents findings from a large-scale field experiment involving 776 Procter & Gamble professionals to study how generative AI reshapes teamwork. The study found that individuals using AI performed as well as traditional two-person teams, highlighting AI’s potential to replicate core aspects of human collaboration. AI also helped break down functional silos between R&D and commercial roles, producing more balanced solutions regardless of professional background. Additionally, the language-based interface of generative AI boosted participants’ emotional responses, suggesting AI can partially fulfill the social and motivational functions of a teammate. These results imply that AI adoption in knowledge work may require organizations to rethink how they structure collaboration and expertise sharing.

• Early methods for studying affective use and emotional well-being on ChatGPT | OpenAI  OpenAI and MIT Media Lab conducted two studies on the emotional impact of ChatGPT use. Analyzing nearly 40 million interactions and a randomized controlled trial with ~1,000 users, they found that emotionally expressive use of ChatGPT is rare and concentrated among a small group of heavy voice-mode users. Text-based interactions showed more emotional cues, while extended daily use—especially of voice—was linked to worse well-being outcomes. Users with strong attachment tendencies were more likely to experience negative effects. The research highlights the need for nuanced understanding of AI's emotional impact and the importance of responsible design and transparent expectations.

• (PDF) Can AI Outsmart Fake News? Detecting Misinformation With AI Models in Real-Time A new study published in Emerging Media evaluates AI’s ability to detect misinformation in real-time using a hybrid approach that combines machine learning, NLP, and deep learning. Researchers tested various models on a balanced dataset of 10,000 claims and 5,000 live posts from the Trump vs. Harris debate. Transformer models like BERT significantly outperformed traditional ML models, with BERT achieving 94.8% accuracy and 93.5% precision. However, real-time deployment is challenged by the models’ heavy compute demands. The study underscores the critical role of high-quality, fact-checked training data and emphasizes ethical safeguards like adversarial testing and model interpretability tools to ensure fairness. 

• 89% of Enterprises GenAI Usage Is Untracked, Posing Security Risks  A new report by LayerX reveals that 89% of generative AI (GenAI) usage within enterprises goes untracked, posing major security risks. Most GenAI interactions occur outside IT oversight, often through personal logins, with only 11.7% of logins using secure corporate Single Sign-On (SSO). Developers are the highest-risk group, frequently pasting proprietary code into GenAI tools. ChatGPT dominates usage (77%), followed by Google Gemini (11%), Claude (5%), and Microsoft Copilot (4%). GenAI browser extensions, installed by 20% of users, can bypass traditional security tools like Secure Web Gateways. To mitigate risks, organizations are advised to implement browser-based security, encryption, data flow monitoring, incident response plans, and regular audits, while adapting their security posture to the AI-powered workplace.

• AI in Government and Public Services Market: Impact of Machine Learning, NLP, and RPA on Public Administration Efficiency  The global AI in Government and Public Services market is projected to grow from $19.2 billion in 2023 to $59.6 billion by 2031, at a CAGR of 15.8%, according to InsightAce Analytic. AI adoption in the public sector is being driven by the need to improve operational efficiency, reduce administrative burdens, and enhance citizen engagement. Key technologies include machine learning, NLP, computer vision, RPA, and expert systems, with applications spanning healthcare, law enforcement, education, and urban planning. North America leads the market due to robust infrastructure and government investment. Challenges include data privacy, high implementation costs, and regulatory complexities. Major players include IBM, Microsoft, Amazon, SAP, Palantir, and Oracle.

• https://arxiv.org/pdf/2503.14499 The paper "Measuring AI Ability to Complete Long Tasks" introduces a novel metric called the "50%-task-completion time horizon" to assess AI systems' capabilities in terms of human performance. This metric represents the duration of tasks that AI models can complete with a 50% success rate, corresponding to the time humans typically take to accomplish the same tasks. To establish this benchmark, the researchers measured the time taken by domain experts to complete tasks from datasets such as RE-Bench, HCAST, and 66 newly introduced shorter tasks. Their findings indicate that current advanced AI models, like Claude 3.7 Sonnet, have a 50% time horizon of approximately 50 minutes. Moreover, the study observes that since 2019, this time horizon has been doubling roughly every seven months, with a potential acceleration noted in 2024. The enhancement in AI models' time horizons is largely attributed to improvements in reliability, adaptability to errors, logical reasoning, and tool utilization. The paper also discusses the limitations of their findings, particularly concerning external validity, and considers the implications of increasing AI autonomy in relation to potentially hazardous capabilities. The authors project that if current trends persist and are applicable to real-world software tasks, AI systems could, within five years, autonomously perform many software tasks that presently require a human a month to complete. Explanation of findings: Measuring AI Ability to Complete Long Tasks - METR 

Concerns

Character.ai launched Parental Insights so parents can see which bots their teens chat with—but don’t worry, they won’t see the conversations, just how much time your kid spent talking to a medieval knight named Steve. Classic April Fools’ twist: transparency without the drama. Anthropic just won a round in a big copyright case over song lyrics. For now, Claude can still hum “Single Ladies”—but only if you ask politely and don’t tell Beyoncé. April’s off to a lyrical start. Australian authors are upset that their books may have helped train Meta’s AI without consent. April Fools? Not quite—more like “April surprise.” The authors want answers, and honestly, they deserve them. Turns out 88% of AI pilots never make it to production. Why? Because saying “let’s do AI” isn’t the same as actually being ready for AI. Just a gentle April reminder: strategy > shiny objects. Researchers are trying to make AI less power-hungry, since some models are basically tiny energy monsters in disguise. Perseus, a clever little tool, can slow GPUs just enough to save energy—like AI putting itself in slow motion to go green. No foolin’.

• Character.ai can now tell parents which bots their kid is talking to CharacterAI has introduced a new “Parental Insights” feature aimed at giving parents more visibility into how their teens use the platform, without disclosing chat content. The optional tool allows teens to send a weekly email report to a parent, detailing their average daily usage and time spent with each bot. This move follows rising concerns and lawsuits over exposure to inappropriate content, particularly involving minors. In response, CharacterAI has implemented safety measures like segregating under-18 users into a more filtered model and emphasizing that bots are not real people. With regulatory scrutiny growing, this is likely just the beginning of the platform's compliance efforts. 

• Anthropic wins early round in music publishers' AI copyright case | Reuters  ​A California federal judge has denied a preliminary injunction sought by Universal Music Group and other music publishers against artificial intelligence company Anthropic. The publishers alleged that Anthropic infringed copyrights by using lyrics from over 500 songs, including works by Beyoncé and the Rolling Stones, to train its AI chatbot, Claude. U.S. District Judge Eumi Lee ruled that the publishers' request was overly broad and failed to demonstrate "irreparable harm" caused by Anthropic's actions. This lawsuit is among several in which copyright holders accuse AI developers of misusing protected content without consent to train AI models. Tech companies like OpenAI, Microsoft, and Meta Platforms argue that their use of such material constitutes "fair use" under U.S. copyright law, a pivotal issue in these legal battles. Judge Lee's decision did not specifically address the fair use defense. The outcome of this case could significantly impact how copyrighted materials are used in AI training, influencing the balance between intellectual property rights and technological innovation.

• ‘No consent’: Australian authors ‘livid’ that Meta may have used their books to train AI  Australian authors are outraged after discovering their books may have been used without consent to train Meta’s AI models via the allegedly pirated LibGen dataset. The dataset includes works by prominent Australian figures like former prime ministers and journalists. Authors such as Holden Sheppard, Tracey Spicer, and Alexandra Heller-Nicholas expressed feelings of violation, exhaustion, and anger, criticizing Meta for profiting off their lifetime work without permission or compensation. The Australian Society of Authors is calling for government action and stronger AI-specific copyright legislation. Meta declined to comment due to ongoing U.S. litigation.

• 88% of AI pilots fail to reach production — but that’s not all on IT | CIO According to new research by IDC and Lenovo, 88% of enterprise AI proof-of-concept (POC) projects fail to reach production. Key reasons include unclear goals, poor data readiness, lack of in-house AI expertise, and weak IT infrastructure. Business pressure, especially from CEOs and boards, is pushing companies to greenlight numerous gen AI pilots—many without strong business cases or sufficient funding. Analysts warn that the bar for gen AI POCs has dropped, and some projects are being padded with AI just to gain approval. While some experts argue that failed POCs still provide valuable lessons, systemic issues in planning, data quality, and ROI tracking remain major barriers to success.

• Can we make AI less power-hungry? These researchers are working on it. - Ars Technica  As demand for AI surges, especially in data centers, energy consumption is spiking — and so are concerns. Power use in U.S. data centers rose from 76 TWh in 2018 to 176 TWh in 2023, driven largely by training and running large language models (LLMs). A single GPT-4 training cycle is estimated to consume 50 GWh, enough to power a small town for a year. Inference now accounts for 60% of an AI model’s lifetime energy use, compared to 40% for training. To tackle this, researchers are turning to model-level optimizations like pruning and quantization, and process-level tools like Perseus, which smartly paces GPUs to reduce waste. Perseus alone can cut training energy by up to 30%. Still, measuring actual energy usage is difficult due to a lack of transparency from companies like OpenAI and Google. Most public figures — like the oft-cited “10x power per AI search” — are based on outdated or unverifiable estimates. Efforts like the ML Energy Initiative’s ML.ENERGY Leaderboard and ZeusMonitor tool aim to benchmark real-world AI energy use. For instance, Meta’s open-source LLaMA 3.1 405B model consumes about 0.93 watt-hours per request, far below speculative estimates for ChatGPT. Despite rapid efficiency gains in GPUs and software, data center electricity use could reach 6–12% of total U.S. power consumption by 2030, and up to one-third in AI-heavy regions like Ireland. Long-term fixes may come from photonic chips or 2D semiconductors, but scaling remains the challenge. Ultimately, researchers argue transparency is key: without real benchmarks and peer-reviewed data, energy optimization is guesswork.

Case Studies 

Healthcare: AI’s now diagnosing, summarizing, and maybe scheduling your colonoscopy. Doctors love it, execs are fighting over who’s in charge, and somewhere an AI is writing your insurance appeal in iambic pentameter. April Fools? Nope, just your new medical assistant with zero bedside manner. Marketing: Gen AI is the overachiever marketers secretly love—churning out ads, insights, and campaign plans before lunch. The twist? It’s your co-creator now. Just don’t let it name your product “MoistBot 3000.” Retail: AI shopping traffic is up 1,200%. Amazon’s bots are recommending you hobbies based on hobbies you didn’t know you had. Foolish? Maybe. Effective? Ask your credit card. Finance: Moody’s bet big on gen AI, proving even the risk police are tired of playing it safe. Meanwhile, 89% of GenAI use in companies is completely untracked. That’s not a joke—unless you’re in security. Engineering: Devs love AI so much they’re letting it write buggy code—and still shipping faster. It’s all fun and games until prod goes down and the AI blames you. Happy April Fools’. The AI’s in charge now.

Healthcare

• Doctor explains how artificial intelligence is already being deployed in medical care | CNN AI is already being used in healthcare through both predictive and generative applications. Predictive AI helps identify high-risk patients (e.g., for sepsis or deterioration) and supports early intervention. It’s also used during procedures like colonoscopies and mammograms to flag abnormalities with higher accuracy. Generative AI, meanwhile, is assisting with tasks like drafting patient notes (ambient AI), summarizing visits, and generating prior authorization letters—reducing administrative burdens and improving clinician efficiency. However, concerns remain around data quality, transparency, and potential misuse by insurance companies. Adoption is growing, but careful validation and regulation are needed to ensure safety and trust.

• Generative AI in healthcare: Current trends and future outlook | McKinsey McKinsey’s latest survey shows that 85% of U.S. healthcare leaders are either exploring or implementing generative AI (gen AI) solutions, with a growing number moving from proof-of-concept to full-scale implementation. Most are pursuing partnerships, particularly with existing vendors and hyperscalers, to build customized solutions rather than developing in-house. Gen AI is primarily being used to boost administrative efficiency, clinical productivity, and patient engagement, with 64% of those with implemented solutions already seeing positive ROI. While challenges such as regulatory complexity and talent gaps remain, healthcare leaders view gen AI as critical to improving operations, reducing costs, and enhancing patient experience, and are adopting governance strategies to ensure responsible deployment.

• Who Should Lead the GenAI Charge? Health Care C-Suites Disagree - Newsweek A new Accenture report reveals significant misalignment among U.S. health care C-suite leaders on who should lead generative AI adoption. While 28% of CEOs see themselves as responsible for redefining roles impacted by GenAI, 80% of other executives believe this should fall to the chief digital or AI officer. The report, based on a survey of 300 executives, highlights fragmented decision-making and lack of clear leadership, which could hinder AI implementation. Experts recommend a multidisciplinary approach led by the CEO to ensure alignment. Missed opportunities were also noted in nonclinical areas like call centers, despite their potential for GenAI-driven efficiency.  

Marketing

• How Should Gen AI Fit into Your Marketing Strategy? Generative AI is transforming marketing by enabling automated content creation, personalized customer interactions, and accelerated market research. A recent HBR article recommends that companies balance automation with human oversight to use gen AI effectively. Salesforce’s “State of Marketing” report, which surveyed 5,000 marketers globally, found that implementing AI is their top priority. Companies like Vanguard and Unilever have already seen measurable results: Vanguard improved LinkedIn ad conversions by 15%, while Unilever cut customer service response times by 90% by incorporating gen AI into their workflows. The article emphasizes aligning AI tools with task complexity and viewing gen AI as a co-creator rather than a full replacement for human input.

My take: Gen AI is not your enemy—it’s your power tool. The marketers who win in this new era won’t be the ones who write every word by hand. They’ll be the ones who know how to use Gen AI to create fresh content, repurpose what already exists, draft campaign plans, run competitive analysis, and gather market insights—all before lunch. That’s not replacing creativity. That’s multiplying it. This isn’t about replacing people. It’s about removing the grunt work so your team can focus on strategy, storytelling, and speed. The smartest marketers I know are using Gen AI as their unfair advantage—not avoiding it out of fear. Use it well, and you don’t just stay in the game—you start playing a whole new one.

Retail and e-Commerce 

• Adobe Analytics: Traffic to U.S. Retail Websites from Generative AI Sources Jumps 1,200 Percent Adobe’s latest analytics report reveals that traffic to U.S. retail websites from generative AI sources has surged by 1,200% since July 2024, driven by chat interfaces assisting consumers with research, recommendations, and personalized shopping tasks. Although AI-driven traffic still trails traditional channels like paid search, it shows 8% higher on-site engagement, 12% more pages viewed, and a 23% lower bounce rate. However, conversion rates remain 9% lower—though markedly improved from a 43% gap in July—signaling AI’s current role in early-stage decision-making. AI-driven desktop traffic accounts for 86% of visits, and conversion rates are highest in electronics and jewelry. Beyond retail, AI-driven traffic to travel sites grew 1,700% and to financial services sites by 1,200%, with AI users spending significantly more time browsing and reporting higher satisfaction. 

• Amazon launches personalized shopping prompts as part of its generative AI push | TechCrunch Amazon has launched a new generative AI feature called “Interests” to make shopping more personalized and conversational. Users can input detailed prompts into the Amazon app or mobile site to receive tailored product recommendations based on preferences, hobbies, and budgets. The tool, powered by large language models, also provides ongoing updates about new items, restocks, and deals matching user interests. Initially available to select U.S. users, Interests joins other AI-powered Amazon tools like Rufus and AI Shopping Guides. The move follows broader trends in e-commerce, with companies like Google also enhancing AI-based shopping features. 

Healthcare

• Here's how AI is transforming finance, according to CFOs | World Economic Forum  According to the World Economic Forum, AI is reshaping the CFO role across industries by enabling automation, data-driven insights, and enhanced risk management. Six CFOs shared their perspectives, emphasizing AI as a strategic imperative for operational agility and long-term competitiveness. While most leaders prioritize partnerships, pilot projects, and measurable ROI, challenges remain, including cybersecurity risks and integration complexity. Key takeaways include the need for disciplined AI investment aligned with business goals, robust ROI frameworks, and strong cybersecurity. AI is seen as central to the future of finance, influencing both near-term pressures and long-term growth strategies.

Software Development

• How Software Engineers Actually Use AI | WIRED WIRED surveyed 730 software engineers to understand how they use AI on the job. The results revealed a deep integration of AI tools like ChatGPT and GitHub Copilot into daily workflows, especially for tasks such as debugging, code generation, and documentation. However, the study also found concerning trends: many developers use AI as a crutch, relying on it even when it produces incorrect or insecure code. While AI boosts speed and productivity, it may be eroding foundational skills and introducing subtle bugs. The takeaway: AI is transforming how developers work—but not without trade-offs in quality, oversight, and skill retention.  

Finance

• How financial institutions can improve their governance of gen AI | McKinsey As gen AI transforms financial services, institutions must overhaul their risk governance frameworks to manage new challenges, including misinformation, legal liability, data misuse, and cybersecurity threats. Traditional AI governance models fall short for gen AI’s complex, multi-step processes and use of both structured and unstructured data. McKinsey recommends a comprehensive approach centered on a gen AI risk scorecard and four categories of risk controls: business, procedural, manual, and automated. The scorecard helps assess customer exposure, financial impact, and model complexity, enabling prioritization of high-risk use cases. Financial institutions should revise oversight structures, splitting responsibility across specialized risk committees, especially for gen AI systems like chatbots that require legal, data, and technical scrutiny. Additionally, institutions need new MRM standards, stronger IP/data-use tracking, and transparency protocols. Manual controls such as human oversight, redaction of sensitive data, and feedback loops remain essential for ethics and compliance. Automated tools can assist in vulnerability detection and real-time response. Procurement teams can also use the scorecard to assess third-party gen AI vendors. Ultimately, the report emphasizes that balancing innovation with robust governance will allow financial institutions to unlock gen AI’s benefits—such as enhanced customer service and decision-making—while managing the risks.

• How a Legacy Financial Institution Went All In on Gen AI Moody’s, a 100+ year-old leader in financial risk assessment, made a bold move in 2023 by aggressively embracing generative AI—a surprising pivot for a company known for its cautious, conservative culture. CEO Rob Fauber led the charge, believing that the risk of standing still in the face of AI disruption was greater than the risk of early adoption. Moody’s began embedding gen AI into core processes like research, ratings, and customer service, framing AI not as a side experiment but as a central transformation strategy. This decision reflects a broader mindset shift: in high-stakes industries, strategic risk-taking is becoming essential for legacy institutions to remain competitive in an AI-driven future.

Learning Center

Anthropic's new paper reveals Claude 3.5 Haiku plans poems in advance, speaks multiple languages with abstract circuits, and even shows clinical reasoning—but don’t worry, it still hallucinates just enough to keep things spicy. April Fools’ twist? Turns out, the bot’s “chain of thought” is sometimes just vibes. Meanwhile, Claude may be growing a brain, but ChatGPT still wins the popularity contest with 4.7 billion monthly visits. Canva came in second—because nothing screams AI revolution like making a birthday flyer. DeepSeek, though, pulled the ultimate April surprise with a 2,026% surge, proving nothing powers growth like mystery and memes. And over at Chatbot Arena—basically the UFC for language models—Gemini 2.5 Pro is leading the fight, followed closely by GPT-4o and Grok. Picture a battle royale where instead of fists, they throw spreadsheets, metaphors, and “helpful suggestions.” The real April Fools’ joke? You can vote, but the bots are probably rating each other behind the scenes anyway. 🤡

Learning

• On the Biology of a Large Language Model  A new study titled “On the Biology of a Large Language Model” explores the inner workings of Claude 3.5 Haiku using circuit tracing to analyze how the model thinks, plans, and reasons. The researchers reveal that Claude 3.5 performs multi-step reasoning, such as determining that Austin is the capital of the state where Dallas is located by first recognizing Dallas is in Texas. In creative tasks like poetry, the model plans ahead by identifying potential rhymes before composing lines. It also processes multiple languages using a mix of language-specific and abstract, language-independent circuits, with a notable shift toward the latter compared to smaller models. In medical scenarios, Claude generates potential diagnoses from symptoms and asks relevant follow-up questions, indicating clinical reasoning capabilities. The model appears to recognize when it lacks sufficient information and has mechanisms to avoid hallucinations—though failures here still lead to occasional inaccuracies. It also includes internal safeguards to reject harmful or unethical requests. Finally, the paper examines how closely the model’s explanations (its “chain of thought”) match its internal computations, showing that while often aligned, some reasoning steps may be more performative than literal. Overall, the study offers a rare look inside a frontier model, underscoring the growing need for interpretability as language models take on increasingly complex and sensitive roles.

Tools and Resources

• Ranked: Most Popular AI Tools by Monthly Site Visits As of January 2025, ChatGPT remains the most visited AI tool globally with 4.7 billion monthly site visits, despite a slight 2.1% drop. No other tool crossed the one billion mark, with Canva coming in second at 887 million. China's DeepSeek saw the largest surge, skyrocketing 2,026% month-over-month to 268 million visits, placing it fourth overall. Other notable tools include Character AI (226M), JanitorAI (200M), and Claude (105M). Google’s Gemini saw a 9.2% dip to 118 million visits, while Microsoft Copilot dropped 17.1% to 101 million. The data, from Aitools dot xyz and visualized by Visual Capitalist, highlights intense global competition, with Chinese platforms rapidly gaining ground as new foundational models enter the market. 

• Chatbot Arena The latest update from Chatbot Arena, a crowdsourced AI benchmarking platform developed by UC Berkeley SkyLab and LMArena, shows Google’s Gemini 2.5 Pro leading the leaderboard, followed closely by OpenAI’s GPT-4o and GPT-4.5, and xAI’s Grok-3 Preview. Gemini 2.5 Pro achieved the highest Arena Score of 1443, indicating strong user preference. Meanwhile, DeepSeek-R1 and DeepSeek-V3 climbed to ranks 7 and 13 respectively, outperforming older Claude models and showing rapid improvement from the open-source player. The platform, which now includes over 2.8 million votes across 220 models, uses the Bradley-Terry model to rank LLMs and highlights the growing competitiveness among providers like Google, OpenAI, xAI, Anthropic, Alibaba, and DeepSeek.

If you enjoyed this newsletter, please comment and share. If you would like to discuss a partnership, or invite me to speak at your company or event, please DM me.