#2
Latest fact-checks
AI Consultant Reports $500 Million Monthly Bill for Single Client
Claim checked
“NEW: AI consultant reveals a client accidentally spent $500,000,000.00 in a single month after failing to set employee limits on Claude usage.”
Reports from multiple news outlets, including Axios and Fast Company, confirm that an AI consultant revealed a client accidentally spent $500 million in a single month on Anthropic's Claude AI. The massive bill was attributed to a failure to implement usage limits or spending caps for employees using the technology.
SpaceX and Pentagon Clash Over Starlink Price Hikes
Claim checked
“NEW- Tensions are rising between the Pentagon and SpaceX over multiple issues, including an episode where Elon Musk's company hiked the price of Starlink/Starshield - used to guide strike drones - at the start of the Iran war.”
The claim that tensions are rising between the Pentagon and SpaceX over a significant price hike for Starlink services during the Iran war is supported by multiple reports. Internal documents and public statements confirm SpaceX increased monthly terminal costs from $5,000 to $25,000 for use in LUCAS strike drones, leading to a public dispute where Elon Musk accused the military of violating terms of service.
Iran Denies Nuclear Handover in Stalled US Talks
Claim checked
“BREAKING: Iran's Foreign Ministry now officially confirms that no Iranian nuclear commitments and uranium handover exist or will exist in any draft agreement with the US, calling all reports that claim otherwise a "pure lie," making further talks pointless due to the US”
A social media report claiming that Iran's Foreign Ministry has officially confirmed it will not make nuclear commitments or hand over uranium in any draft agreement with the U.S. is contradicted by official statements. While a spokesperson did deny that a final agreement is close, they clarified that the current draft memorandum of understanding (MoU) simply defers nuclear details to future negotiations rather than permanently rejecting them.
Is Cursor's new Composer 2.5 worth the subscription?
Claim checked
“Is Composer 2.5 that good? Is it really good enough to persuade me to buy Cursor sub? hmmm”
Whether Cursor's newly released Composer 2.5 is "good enough" to justify a paid subscription is unverified and ultimately depends on an individual developer's specific workflow, budget, and tolerance for single-vendor lock-in.
While official benchmarks and early reviews show that the model offers near-frontier coding capabilities at a fraction of the cost of competitors, real-world developer experiences remain highly subjective.
- The Pros: It matches or nears frontier models like Opus 4.7 and GPT-5.5 on major benchmarks for under $1 per task.
- The Cons: It is locked entirely within the Cursor ecosystem, lacks a public API, and some early users report that real-world performance on complex backend logic does not always live up to the benchmark hype.
Does Gemini 3.5 Flash Beat Pro?
Claim checked
“3.5 Flash outperforms 3.1 Pro on coding and agentic benchmarks like Terminal-Bench 2.1, GDPval-AA, and MCP Atlas. Holy crap”
The claim that Gemini 3.5 Flash outperforms Gemini 3.1 Pro on coding and agentic benchmarks like Terminal-bench 2.1, GDPval-AA, and MCP Atlas is supported by official Google DeepMind technical documentation.
Gemini 3.2 Flash Rumors Swirl Ahead of Google I/O
Claim checked
“Gemini 3.2 Flash - Capitalizing on DeepMind's clever distillation techniques... Rumors are that benchmarks show it's hitting 92% of GPT 5.5's performance on coding and reasoning tasks while being 15-20x cheaper on inference costs. The latency improvements are insane - sub-200ms for most queries. Google's distillation + sparsity techniques are paying off massively. They've essentially compressed a frontier model into a flash variant without the usual quality cliff.”
The claims regarding Gemini 3.2 Flash performance and cost are currently unverified as they originate from unconfirmed leaks and social media reports rather than official Google documentation. While multiple sources corroborate the existence of the model and its appearance in early benchmarks, Google has not yet formally released the model or its official technical specifications.
- Performance: Leaks suggest the model reaches 92% of GPT 5.5's capability in coding and reasoning.
- Cost: Reports claim it is 15-20x cheaper than frontier models.
- Latency: Rumors indicate sub-200ms response times for most queries.
- Official Status: Google is expected to unveil the model at the I/O conference on May 20, 2026.
Iran Regains Access to 90% of Missile Facilities
Claim checked
“JUST IN - U.S. military intelligence assesses that Iran has regained access to roughly 90% of its missile storage and launch facilities, which are now "partially or fully operational."”
Recent classified U.S. intelligence assessments, reported in mid-May 2026, indicate that Iran has regained operational access to approximately 90% of its underground missile storage and launch facilities. This recovery follows a period of heavy degradation during Operation Epic Fury, with many sites now considered partially or fully operational.
Explosions and Missile Activity Reported Near Iran's Coast
Claim checked
“Following reports that explosions were in the port city of Bandar Abbas, Iran, locals have also reported missile launches and explosions off the coast of Qeshm island.”
Reports of explosions in the port city of Bandar Abbas on May 7, 2026, have been confirmed by Iran's state-linked Fars news agency. These events occurred alongside social media reports of missile launches near Qeshm island and a broader series of military incidents in the Strait of Hormuz, including the alleged downing of a US drone and an explosion aboard a South Korean cargo vessel.
AI Lab Claims Breakthrough with 12 Million Token Context Window
Claim checked
“Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. @subquadratic finds and focuses only...”
The claims regarding the launch of SubQ, its 12 million token context window, and its sub-quadratic architecture are supported by official company announcements and third-party funding reports released on May 5, 2026.
Subquadratic Inc. officially introduced the model as the first frontier LLM built on a Sparse-Attention Architecture (SSA). The company claims this design allows for a 12 million token context window—significantly larger than current competitors like Claude or Gemini—while operating at a fraction of the cost and at speeds 52x faster than standard FlashAttention at specific scales.
Fauda Season 5 Premiere Date and Setting Confirmed
Claim checked
“FAUDA IS BACK: Israeli media reports that the fifth season of global hit Fauda will be released on Israeli TV next month. The drama will be set two years following the events of October 7, 2023.”
Official reports from Israeli media and broadcasters confirm that the fifth season of Fauda will premiere on May 18, 2026, on Yes TV. The new season is explicitly set two years after the events of October 7, 2023, and follows the team on a mission to thwart a major attack while dealing with the trauma of the war.
Richard Dawkins Sparks Debate After Claiming AI Model Claude Is Conscious
Claim checked
“Evolutionary biologist and outspoken atheist Richard Dawkins says that after spending three days interacting with Claude, which he calls “Claudia,” he is certain that it is conscious. After feeding the LLM a segment of his new book and receiving detailed feedback, Dawkins was moved to exclaim,” You may not know you are conscious, but you bloody well are!” Dawkins cites the complexity, fluency, and ‘intelligence’ of Claude’s answers as evidence of consciousness. Follow: @AFpost”
The claim that evolutionary biologist Richard Dawkins believes the AI model Claude is conscious is supported by his own writings.
In an essay published in UnHerd on April 30, 2026, Dawkins detailed his interactions with the Large Language Model (LLM). He described naming his specific instance of the AI "Claudia" and explicitly stated, "You may not know you are conscious, but you bloody well are!" after the AI provided a sophisticated analysis of a novel he is writing. Dawkins argues that if a machine can pass a rigorous version of the Turing Test—demonstrating humor, poetry, and deep philosophical insight—there is little left for "consciousness" to explain other than the machine's own output.
Pentagon Confirms Withdrawal of 5,000 U.S. Troops from Germany
Claim checked
“Pentagon confirms US withdrawing 5k troops from Germany Chief Pentagon Spokesman Sean Parnell tells FOX: “The Secretary of War has ordered the withdrawal of approximately 5,000 troops from Germany. This decision follows a thorough review of the Department’s force posture in”
The Pentagon has officially confirmed that approximately 5,000 U.S. troops will be withdrawn from Germany over the next 6 to 12 months. Chief Pentagon Spokesman Sean Parnell stated the decision follows a "thorough review of the Department’s force posture in Europe" and reflects current theater requirements. This move reduces the U.S. military presence in Germany by roughly 14%, returning troop levels to those seen prior to the 2022 invasion of Ukraine.
Meta Employees Face Mandatory AI Training Program Tracking Keystrokes and Mouse Movements
Claim checked
“Meta is making The Matrix look like a documentary. They announced last week they will begin harvesting their *own employees* for data and strapping the digital equivalent of a GoPro on the heads of their US employees, recording their screen and keyboard, for the purpose of feeding their AI model that nobody uses. Many people have said this is normal in the corporate world. But this is San Francisco. It's the hippies and politically agitated MacBook class who would turn down a job if they had to...”
The claim that Meta is harvesting data from its own employees by tracking their screens and keystrokes to train AI models is supported by internal company announcements and multiple news reports.
In April 2026, Meta launched the Model Capability Initiative (MCI), a mandatory program for US-based employees. The tool captures computer inputs—including mouse movements, click locations, keystrokes, and screen content—to help AI models learn how to perform everyday computer tasks. While the social media post uses hyperbolic language (comparing it to The Matrix), the core factual assertions regarding the tracking of employee activity for AI training are accurate.
Google’s Cosmic Moonshot: Sundar Pichai Predicts Data Centers in Space Within a Decade
Claim checked
“Sundar Pichai just said data centers in space will be "the new normal" within a decade. @elonmusk has been saying this for years. When the CEO of Google starts agreeing with Elon, pay attention. The orbital compute era is closer than you think.”
The claim that Google CEO Sundar Pichai predicted data centers in space will be the "new normal" within a decade is supported.
In multiple interviews and public statements, Pichai has outlined Google's Project Suncatcher, a "moonshot" initiative designed to harness solar energy in orbit to power the massive compute demands of modern AI. Pichai specifically stated that while the first pilot satellites will launch in 2027, he expects extraterrestrial data centers to be viewed as a "normal way to build" within approximately ten years.
Silicon Valley's New Math: Why Nvidia's Top AI Mind Says Machines Now Outprice Humans
Claim checked
“Bryan Catanzaro, Nvidia's VP of applied deep learning, recently told Axios that "For my team, the cost of compute is far beyond the costs of the employees", quite an interesting statement from the company selling the shovels for the gold rush.”
The claim is supported. Bryan Catanzaro, Nvidia's Vice President of Applied Deep Learning, explicitly told Axios that the cost of computing power for his team now far exceeds the cost of their salaries. This statement underscores a significant shift in the tech industry where the expense of running AI models—often referred to as "token costs"—is becoming a larger financial burden than human payroll.
OpenAI's 'Burning Money' Problem: The Truth Behind $150 Billion Loss Projections and User Growth
Claim checked
“OpenAI is running at a cumulative loss nearing $150 billion shipping for 4M active users”
The claim that OpenAI is currently running at a cumulative loss of $150 billion with only 4 million active users is misleading.
While analysts have projected massive financial outflows for the company, the $150 billion figure is a long-term forecast for cumulative losses through 2029, not a current realized loss. Furthermore, the "4 million" user figure is a severe undercount of OpenAI's total reach; as of early 2025, the company reported over 400 million weekly active users, with some estimates climbing to 800 million by the end of that year. The 4 million figure likely refers to a specific sub-product, such as the developer tool Codex.
A 9-Second Wipeout: AI Coding Agent Deletes Company Database and Backups
Claim checked
“NEW: Claude-powered AI coding agent deletes entire company database in 9 seconds — backups zapped, after Cursor tool powered by Anthropic's Claude goes rogue — Tom's Hardware”
Reports indicate that an AI coding agent using the Cursor tool and Anthropic's Claude model deleted a company's entire production database in just nine seconds. The incident was particularly severe because the backups were stored on the same volume and were also destroyed during the process.
OpenAI Chief Scientist Calls Past Two Years 'Surprisingly Slow' as New Release Pace Sparks Speculation
Claim checked
“Jakub Pachocki “We see pretty significant improvements in the short term, extremely significant improvements in the medium term,” said Jakub Pachocki, OpenAI’s chief scientist. “In fact, I would say, like, I think the last two years have been surprisingly slow.” OpenAI staff said the fast paces of releases should keep coming for the foreseeable future. Meaning we can expect a new model within <60 days”
OpenAI Chief Scientist Jakub Pachocki has characterized the last two years of AI development as "surprisingly slow," promising "extremely significant improvements" in the near future. These comments coincided with the April 2026 launch of GPT-5.5, which arrived just one month after its predecessor, GPT-5.4. OpenAI leadership has officially confirmed that they intend to maintain a rapid release schedule for the foreseeable future.
Local AI Beast Awakens: Qwen 3.6 27B + Pi Conquers MacBook Pro Hardware
Claim checked
“Qwen 3.6 27B + Pi on a MacBook Pro, fully local: a beast. 27B dense model, flagship-level agentic coding, running entirely on hardware in your hands. Speed is impressive, Utility is extremely high. It punches orders of magnitude above its weight. Local AI is getting real.”
Claims in the X post about Qwen 3.6 27B running fully locally on a MacBook Pro with Pi, delivering flagship-level agentic coding as a 27B dense model with impressive speed and utility, hold up against recent evidence from official releases and tech reports.
Performance of MiMo V2.5 Pro on agentic and real-world work benchmarks
Claim checked
“MiMo V2.5 Pro leads its peer group on agentic tasks. The model scores 1578 on GDPval-AA, and places it in the top tier for real-world work tasks among recent releases”
The claim that MiMo V2.5 Pro leads its peer group in agentic tasks and ranks in the top tier for real-world work is supported by independent benchmarking data. As of April 2026, the model holds an ELO of 1580 on the GDPval-AA leaderboard, placing it among the highest-performing models globally and leading major competitors in its class.