[AI DAILY NEWS RUNDOWN] Musk Loses to OpenAI, the Data Center Revolt, and Open-Weight GPT (May 18 2026)

Two days before the company released GPT-5 in August 2025, it quietly announced two open models, GPT-OSS-120B and GPT-OSS-20B. In technical terms, these are Mixture-of-Experts (MoE) models that use chain-of-thought reasoning. That makes them very good at math, programming, and research tasks. The 120B and 20B designations refer to the number of parameters in each model.

Despite flying under the radar, the GPT-OSS models are generally considered state-of-the-art open-weight models and operate under the permissive Apache 2.0 license, which gives developers and enterprises lots of leeway to customize them.

OpenAI told The Deep View that the company released these models because, due to regulatory and security requirements, some organizations must run workloads locally rather than in the cloud. That’s where open models have to play a critical role, as they can run in the company’s own hardware and data center. In practice, OpenAI reported that many of these organizations operate a hybrid environment, with some workloads running on GPT-OSS and others running on OpenAI’s latest frontier models in the cloud.

The two OpenAI models serve different purposes:

• GPT-OSS-20B: The smaller of the two models is designed to run on local hardware, such as a laptop. In fact, it can run in 16 gigabytes of memory and works on nearly any laptop less than five years old. It can save you money in token costs by running locally while running very fast. It’s also more secure and lets you run without an internet connection. It’s the more popular of the two on Hugging Face.

• GPT-OSS-120B: The larger of the two models offers configurable levels of reasoning and is powerful enough to rival proprietary models such as OpenAI’s o4-mini. But it’s efficient enough to run on a single 80GB GPU in a data center. It’s considered incredibly fast and compute-efficient. It’s so efficient that it can even run on a MacBook Pro that has enough memory.

When The Deep View met with the team at OpenAI behind the GPT-OSS models and asked why the company’s open models don’t get more attention, the team shared a key insight. The broader open-model community tends to favor models that are easy to fine-tune with supervised learning. GPT-OSS behaves more like an OpenAI reasoning model with web search, which makes it more capable but harder to customize. It rewards reinforcement learning instead.

Also, not to be forgotten, three months after releasing these two open models, OpenAI released two more aimed at safety: GPT-OSS-Safeguard-120B and GPT-OSS-Safeguard-20B. These models let you bring your own safety policies and enterprise guardrails, enabling auditable safety behavior.

Alexa+ now generates custom AI podcasts LINK

• Amazon rolled out a feature called “Alexa Podcasts” on Monday, letting Alexa+ users in the U.S. ask the assistant to create a podcast episode about any topic they want, ready in minutes.

• Users skip uploading documents or writing scripts, and instead Alexa+ researches the topic, drafts an overview, and lets people adjust the length, tone, and focus before AI-generated host voices narrate the episode.

• Finished episodes trigger a notification on Echo Show devices and the Alexa app, where they are saved under “Music” and “More,” and pull real-time information from partners like the Associated Press, Reuters, and The Washington Post.

Artist shines mirror on AI anger with viral Monet post

Image source: Images 2.0 / The Rundown / @SHL0MS on X

The Rundown: Artist SHL0MS sparked a frenzy on X over a post that presented a real image of a painting from famed impressionist Claude Monet as AI-generated, sparking critiques of the work and shining a mirror on anti-AI bias seen in the creative world.

The details:

• SHL0MS posted that he generated the image in the style of Monet, asking users to describe why the “AI image” is inferior in as much detail as possible.

• The post received thousands of responses calling the image ‘emotionless’, ‘slop’, and critiquing specific features like depth, reflections, and composition.

• The image was an image of a real Monet, identified as from his Water Lilies collection from around 1915.

• The saga aligns with 2024 research, with Norwegian researchers finding people actually prefer AI art but show a clear negative bias against it.

Why it matters: This Monet-gate was one painting, but the reflex it exposed runs through the entire creative world right now. For a growing crowd, the word ‘AI’ alone triggers backlash regardless of context — and that knee-jerk hostility is only growing as the tech changes the world around us and embeds deeper into everyday life.

ChatGPT starts connecting to your money

Image source: OpenAI

The Rundown: OpenAI released a new personal finance experience inside ChatGPT, partnering with Plaid to connect users’ financial institutions and give the chatbot real-time access to spending, investments, and bills for personalized financial insights.

The details:

• Users can connect via Plaid across Chase, Schwab, Robinhood, and 12,000+ institutions, with a dashboard tracking spending, portfolio, and upcoming bills.

• Intuit support is planned next for flows like tax estimates, credit-card approval odds, and connecting users to live experts.

• ChatGPT can analyze connected data, but cannot move money, pay bills, make trades, or file taxes (yet).

• The feature preview starts with U.S.-based Pro accounts, accessible via a new ‘finance’ sidebar or by tagging @finance directly in chats.

Why it matters: Financial guidance has been locked behind pricy advisors or clunky apps for a decade, and having a complete view of a user’s finances as context can unlock a seriously powerful experience. The biggest hurdle might be getting users to trust AI with that information, which is where Plaid comes in as the secure layer.

Alibaba’s AI bets soar:

Chinese tech giant Alibaba reported 3.2M robotaxi rides in Q1 2026, up from 1.4M in the year-ago quarter. Elsewhere in its AI results, the company’s GPU cloud saw its growth rate rise from 128% in Q3 2025 to 143% in Q4 2025 to a staggering 184% in the most recent period. While much attention is paid to American AI labs, don’t forget just how big — and intelligent — the world is.

Apple’s revamped Siri to auto-delete chats LINK

• Apple’s reworked Siri app, set to arrive with iOS 27, iPadOS 27, and macOS 27, will automatically delete chat conversations by default as part of a privacy-first design the company plans to highlight at WWDC.

• Users will reportedly be able to keep Siri chats for 30 days, one year, or forever, matching the options already found in Apple’s Messages app, with tighter limits on what memory can persist and for how long.

• Mark Gurman reports that Apple will argue its approach differs from rivals that train models on user interactions and cloud-stored histories, and the new Siri is expected to launch with a “beta” tag and a toggle to exit it.

Torvalds says AI bug reports overwhelm Linux security list LINK

• Linus Torvalds says the Linux security mailing list has become “almost entirely unmanageable” because researchers are using AI tools to hunt for bugs and then flooding the list with duplicate reports of the same issues.

• Torvalds delivered the complaint in his weekly state of the kernel post alongside release candidate four for Linux 7.1, noting maintainers waste time forwarding messages or replying that a bug was already fixed weeks or months ago.

• He argued AI-detected bugs are by definition not secret, so handling them on a private list only worsens duplication since reporters can’t see each other’s submissions, and he asked people to read the docs and write a patch.

Google researchers queue for TPU capacity LINK

• Google’s own AI researchers, including teams at DeepMind, are now waiting in line for TPU compute access because the chips are being sold in bulk to outside customers like Anthropic and Meta, Bloomberg reported Monday.

• The squeeze stems from Google’s $40bn Anthropic deal covering five gigawatts of TPU capacity and up to one million Ironwood chips over five years, plus a separate Broadcom-mediated 3.5GW supply line starting in 2027.

• Bloomberg notes researchers including DeepMind’s Ioannis Antonoglou have left for startups as internal compute gets harder to secure, with Oren Etzioni saying access is rationed by managerial seniority rather than the economics applied to external contracts.

xAI fails to pay employees promised $420 for their tax data

• Two months after xAI asked employees to share their personal US tax returns to train Grok in exchange for a $420 payment per submission, none of the promised money has actually been handed out, Bloomberg reports.

• The collection drive was timed to the April 15 tax deadline, framed internally as a way to feed Grok real, complex filings that xAI cannot license at scale or scrape from the open web.

• Employee returns contain salary, dependents, addresses, financial-account positions and Social Security numbers, raising questions about data-handling promises, while the total bill for paying every submitter sits in the low six figures.

AI Lowers Barrier for Threat Actors

A recent Cybersecurity Insider survey found 47.1% of respondents believe AI is creating more capable cybercriminals, while 29.4% said it is lowering the barrier for inexperienced attackers.

Threat actors are already using AI for phishing and malware development as attacks become more scalable, convincing, and difficult to detect.

Source: ChatGPT

At the same time, many of you still believe successful cybercrime operations require technical expertise, operational security, and strategic planning beyond what AI alone can provide.

Strengthen behavioral detection, phishing-resistant MFA, and employee awareness training to help defend against personalized AI-assisted attacks.

OpenAI Confirms Developer Device Compromise

OpenAI confirmed two employee devices were compromised in the TanStack supply chain attack, highlighting growing threats targeting developer ecosystems and CI/CD infrastructure.

The broader Mini Shai-Hulud campaign compromised hundreds of npm and PyPI packages using stolen CI/CD credentials and malicious updates to steal developer and cloud credentials while maintaining persistence through modified developer tooling.

The incident highlights growing software supply chain risks tied to open-source dependencies and automated deployment workflows.

Strengthen CI/CD security with least-privilege access, short-lived credentials, dependency monitoring, and regular supply chain attack simulations.

AI hasn’t overtaken human writers online

By Megan Morrone

Data: Graphite.io; Note: Based on an average of three AI-detector tools sampling URLs from Common Crawl; Chart: Megan Morrone/Axios

The flood of AI-generated writing unleashed by ChatGPT appears to have leveled off — a sign that AI content hasn’t overtaken the web after all.

The big picture: The share of online news articles, blog posts and listicles that are primarily AI-generated has held near 50% for more than a year, according to a new analysis from digital marketing agency Graphite.

• The plateau indicates that the feared takeover of human online writing by AI hasn’t materialized — at least not yet.

Why it matters: Researchers who’ve studied the spread of AI-written articles warn that once models start training on that content, the internet could become a massive feedback loop of low-quality, machine-generated content.

• “These models are smart because of all the information we put on the web that was created without these models,” Dan Klein, a UC Berkeley professor and AI model CTO, tells Axios. “If we stop creating knowledge that is independent of these models, what’s going to fuel that?”

Keep reading.

3 ways to write AI prompts to get better results

AI chatbots arrived with a simple promise: talk naturally, get answers. But users quickly discovered that prompting AI to get the results they actually wanted was far more complicated.

That skill gap became so valuable that an entirely new role emerged, the prompt engineer, commanding surprisingly generous salaries. As the technology matured and more people learned the basics, that specific title faded. But the skill itself didn’t become less important — it became more so.

When anyone can generate a generic prompt, those who know how to extract real value from AI are the ones with the competitive edge.

“The users who understand its nuances, abilities, and limitations will be the ones who unlock real value from it, while the rest will be stuck with generic outputs and false confidence in them,” Eric So, distinguished professor of global economics, behavioral science and management at MIT Sloan, told The Deep View.

There are proven practices for getting the best results from AI, and after speaking with the experts, we’ve rounded up the most useful ones below.

1. Include lots of context: AI doesn’t know you and your preferences the way a human does, so you need to frontload as much information as you can.

2. Include examples: Sometimes, AI can learn to complete a task more effectively by mimicking an example than by following instructions.

3. Iterate: “Don’t try to get it perfect on the first try. The best results usually come from an iterative back-and-forth — ask, see what you get, then refine,” Neel Joshi, director of product management for Gemini Apps, told The Deep View.

If you’ve tried your hand at including as much context as possible to steer the model, but you’re still not getting the results you want, you can also have the model interview you to gather the information it needs to fill in the gaps.

“You can provide all the information up front, or you can use a flipped interaction where the model interviews you interactively to elicit the information it actually needs,” Jules White, Ph.D., professor of computer science at Vanderbilt University. “That second approach is powerful because the AI can adapt its questions dynamically and uncover gaps or requirements you may not have anticipated yourself.”

Situational Awareness files hotly anticipated 13F (TBPN)

Situational Awareness’s latest 13F is now public. Leopold Aschenbrenner, the hedge fund’s Chief Investment Officer, is known for making extremely successful investments based on his core assumption that frontier AI will continue to improve at 0.5 OOMs (orders of magnitude) per year, which translates into a thesis that AI will create unprecedented demand for compute and its associated bottlenecks.

There’s a ton of discourse around it on the timeline, mostly because the filing shows he’s made some massive puts across the semiconductor sector (e.g. ~$2B on SMH, the VanEck Semiconductor ETF). But the filing is hard to interpret cleanly. A 13F is only a snapshot of holdings as of March 31, 2026, meaning these positions were in place during the early phase of the Iran war. And while 13Fs do disclose put/call positions, they don’t disclose the strike prices, expirations, premiums paid, hedge ratios, short positions, swaps, or whether the options are part of broader structures.

This tweet by fejua sums up this dynamic nicely. — Brandon

AI Backlash Continues (TBPN)

Big pushback on AI data centers across the board. It’s both a left- and right-wing issue now (as predicted during Saagar Enjeti’s TBPN appearance last year). The latest debate is over a huge data center in Utah that’s being championed by Shark Tank’s Mr. Wonderful (Kevin O’Leary). He’s a bit of an over-the-top caricature of a businessman, recently seen sporting not one, but two expensive watches at the Oscars (a Cartier Crash Skeleton and a Ruby Rolex Daytona). It makes him a bit of a soft target for anyone who wants to paint data center construction as maybe not in the best interest of average Americans. But the Utah data center project actually seems pretty by the book according to current plans. It’s in a remote area, uses its own power and water, and doesn’t seem to disrupt any local communities. Quick Thoughts has a good breakdown. These points are going to be hard to breakthrough though, because AI is just so deeply unpopular. Just watch this video of Eric Schmidt getting booed while giving a commencement speech.

There are like 10 different problems to solve in order to turn the tide here. AI has to make more cool things, everyone is vibecoding 24/7 and leaving their MacBook Pros open so they can work while they are at parties, but we aren’t seeing delightful new consumer apps pop up in any meaningful way. Then AI also has to create jobs, or at least the job market has to independently strengthen. On the data center side, addressing environmental impact (probably real), water use (mostly fake), noise (solvable), and a bunch of other issues is not going to happen magically. Ben Thompson had a good analysis of “Data Center Discontent” today in Stratechery and shared a pretty simple solution:

Just to put some numbers on this, the data center up the road was expected to be 1.6 GW, which could generate around $3 billion in annual operator revenue. DeForest, the village it was to be built in, has around 11,500 people. You could pay every person in DeForest $10,000 a year for 3.8% of annual revenue grossed by the data center — I bet that proposal would have been approved, and I bet that the operator could very easily pass those costs on to the actual data center users (it also highlights how relatively pathetic QTS’s $50 million commitment was).

All the big labs like to say that “things get weird” in the AI era, and I wouldn’t be surprised if we see lots of weird twists and turns in the data center buildout and accompanying pushback. — John

Chinese fast-fashion company Shein buys Everlane (TBPN)

Puck News has reported that Shein, the Chinese company known for pioneering algorithmically driven fashion trends on the back of extremely inexpensive clothing, has acquired Everlane in a deal valued at $100M. The transaction follows years of financial strain at Everlane during which the company took on about $90M in debt. Lots of takes on the timeline about this, one of the most salient throughlines being that Everlane — which built its brand on transparency and ethical operations — essentially sold out to one of the least scrupulous companies in the fashion category (at least along the ethical dimension). — Brandon

What’s the root cause of falling birth rates? (TBPN)

The Financial Times has a great deep dive in The Big Read about falling birth rates. The screenshot that went viral puts most of the blame on the rise of smartphones. The actual article covers lots of different factors and highlights rising housing costs as a big piece of the puzzle, but the data around smartphone adoption is pretty convincing. Ross Douthat and others shared that, outside of the temporary baby boom in the mid-20th century, birth rates have been falling steadily since 1800. There’s an interesting debate around what happens if you try to go a layer deeper into the question. What apps specifically are causing a birth rate decline (if you buy that conclusion)? Is it dating apps that create unrealistic expectations? Social media sites that flood you with videos of lives just out of reach, but worth delaying family formation for a shot at? Video games? Gambling? Etc etc. The smartphone is such a bundle of features (famously a phone, iPod, and an “Internet Communicator”). I don’t think it’s music on demand that’s causing people to put off having kids, but is there a version of a smartphone, or a series of apps that cause stronger relationships to develop? Maybe the next study should be comparing screen time or specific content that’s served to adults who have kids vs. those who don’t. All this is interesting because as fast as the AI takeoff is going, many of the social and economic changes that we actually see showing up in the world are just taking shape after two decades of smartphones and social media diffusion. It does give a lot of credence to the “we should be very thoughtful about how technology rolls out earlier than ever” arguments. — John

What Else Happened in AI on May 18th 2026?

OAI’s Greg Brockman is reportedly moving to product strategy, with Thibault Sottiaux now leading core product and platform, and Nick Turley shifting to enterprise products.

Anthropic formed a $200M partnership with the Gates Foundation to deploy Claude in vaccine screening, disease forecasting, and K-12 tutoring in developing nations.

OpenAI reportedly acquired Weights.gg in January, a voice-cloning social network with celebrity replicas, with six staff members joining the company.

Pope Leo XIV established a Vatican AI advisory body and signed his first encyclical, which is expected to frame the technology as this generation’s Industrial Revolution.

Poetiq released new research showing its self-building Meta-System can create and improve on its own harness, and achieve SOTA scores on a top coding benchmark.

A Gallup survey found that 70% of Americans oppose data centers being built nearby, with the infrastructure polling as less popular than local nuclear power plants.

NYT: Jury Rejects Musk’s Claim Against OpenAI

Situational Awareness releases 13F

FT: Why birth rates are falling everywhere all at once

Stratechery: Data Center Discontent, Understanding the Opposition, Fixing the Problem

Nic Carter: AI is not hiking your electricity bill – yet

Non-Public Policy: In Defense of the Data Center

NBC: A trade boost and a new relationship status: U.S. touts modest China deals after summit’s good vibes

FT: In America, wealth taxes mean wealth flight

FT: The fate of OpenAI’s $1tn IPO will be decided in an Oakland jury room

WSJ: Amazon Web Services CEO Pushes Back on AI Job Apocalypse Warnings

WSJ: Apple Is Making Hit Products and High Profits From Imperfect Chips

Eric Seufert: Components of AI-enabled advertising

WSJ: The Tech Bros Are Going to Etiquette School

Claire Zau joins Lightspeed as a partner

Mark Cuban proposes taxing tokens