Weekly AI Roundup: What Happened in AI This Past Week (May 2–8, 2026)

Pull up a chair. The kettle is on. If you have been hearing your nephew or your colleague say “GPT this” and “Gemini that” all week and nodding politely without really following along, this post is for you. The last seven days have been extraordinary — not because of one single moment, but because AI moved on five fronts at once, and some of what happened this week will still matter years from now.

I have split this week into two parts: five big stories from the global stage, then five more that hit closer to home in Bangladesh. Whether you are a freelancer in Mirpur, a researcher in BUET, or a student sitting in a university canteen, something here will land.

5 Big International Developments

1. ChatGPT quietly became a lot smarter as its new default

On Tuesday, May 5, OpenAI swapped in GPT-5.5 Instant as the default model for all ChatGPT users — the engine that runs every time you open a new conversation. This is not a headline model aimed at power users. It is the everyday ChatGPT that billions of people use for everything from drafting emails to looking up recipes.

The headline number from OpenAI: GPT-5.5 Instant produces 52.5% fewer hallucinated claims than its predecessor GPT-5.3 Instant on high-stakes questions in law, medicine, and finance. Put plainly, it is significantly less likely to confidently make something up. Accuracy on the GPQA benchmark jumped from 78.5% to 85.6%, and on maths competition problems from 65.4% to 81.2%.

OpenAI also introduced “memory sources” this week — when ChatGPT personalises a response using your saved chats or connected Gmail, it now shows you exactly which information it used, so you can correct or delete it. This is a small thing that quietly fixes a creepy problem: AI using your personal data invisibly.

For most people reading this, the practical takeaway is simple. ChatGPT just got noticeably better at not lying to you. That matters more than any benchmark.

2. Anthropic crossed OpenAI in revenue — for the first time

This is the headline that shook Silicon Valley’s sense of the natural order. Anthropic, the company behind Claude, reported $30 billion in annualised revenue this week, reportedly passing OpenAI for the first time. OpenAI, by comparison, is targeting roughly the same figure but reportedly missed several internal milestones this quarter.

To put Anthropic’s trajectory in perspective: the company’s revenue grew thirty times in fifteen months. That kind of pace does not happen by accident. It happened because Claude has been winning a string of coding and enterprise benchmarks. Claude Opus 4.7, released a fortnight ago, scored 87.6% on SWE-bench Verified — the standard test for resolving real-world software bugs — compared to 74.9% for the equivalent OpenAI model. When your AI is better at writing and fixing code, companies pay for it.

Anthropic followed the revenue news with a concrete product push on May 4-5: ten ready-to-run AI agent templates for financial services, covering pitchbook building, KYC screening, general ledger reconciliation, month-end close, and valuation review. These templates plug directly into Microsoft Excel, PowerPoint, and Word. The same day, Anthropic announced a joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman — $1.5 billion of enterprise services capacity aimed at deploying Claude inside the biggest financial institutions in the world.

The era of a single company running away with the AI industry may already be over.

3. The Pentagon drew a line — and Anthropic was on the wrong side of it

This is the story that felt most like a thriller. The US Department of Defense this week formalised classified AI deployment agreements with eight companies: OpenAI, Google, Microsoft, Amazon Web Services, NVIDIA, SpaceX, Oracle, and Reflection. These deals give the Pentagon access to frontier AI for any lawful government purpose, including classified workloads.

Anthropic was not on the list. The Pentagon had previously designated Anthropic a “supply-chain risk” — a designation usually reserved for foreign adversaries — because Anthropic refused to allow Claude to be used for fully autonomous weapons systems and mass domestic surveillance. It is the first time an American AI company has been locked out of US government contracts specifically because of its own ethical rules.

There is a moral complexity here worth sitting with. Anthropic’s refusal may or may not be the right call strategically. But the fact that a company can be punished commercially for having safety guardrails sends a signal to every AI lab watching: principles cost money.

Meanwhile, Google — which was already the largest private backer of Anthropic through a $40 billion investment announced the previous week — signed the Pentagon deal. One hand funds the safety-conscious lab; the other hand takes the government contract the safety-conscious lab refused. That is not hypocrisy so much as industrial-scale hedging.

4. Meta launched Muse Spark and fired 8,000 people on the same day

Meta had a complicated week. On one hand, the company released Muse Spark, its first AI model from the newly formed Meta Superintelligence Labs, run by Alexandr Wang, former CEO of Scale AI. Muse Spark is the model that was codenamed “Avocado” during a famously troubled development process. It is competitive with OpenAI and Anthropic on multimodal understanding and health-related tasks, though Meta acknowledged it still trails on coding.

The bigger strategic signal: Muse Spark is closed-source, a reversal from Meta’s open Llama strategy. Mark Zuckerberg reportedly grew frustrated with Llama’s pace and the widening gap with frontier models, and has bet $14.3 billion on Wang’s new approach. If Muse Spark works, Meta becomes a genuine fourth competitor at the frontier. If it does not, the company has abandoned the goodwill it built with the open-source developer community for nothing.

On the other hand, Meta also announced it will cut 8,000 jobs on May 20, roughly 10% of its workforce, as it shifts budget toward AI infrastructure. Meta’s projected AI capital expenditure for 2026 is $115 to $135 billion. The company’s Q1 revenue came in at $56.3 billion, up 33% year on year — and the stock still fell 7% because investors are worried about the scale of AI spending relative to returns. The gap between “impressive AI demos” and “reliable AI profits” remains the industry’s central anxiety.

5. OpenAI opened ChatGPT to advertisers

This one crept up quietly alongside the GPT-5.5 Instant launch, but its long-term implications are significant. On May 5, OpenAI announced new ways for brands to buy ads inside ChatGPT. It is framed carefully — sponsored content that appears when relevant — but the direction is unmistakable. The world’s most widely used AI assistant is becoming an advertising platform.

Anthropic explicitly said the same week that Claude will not carry ads, citing incompatible incentives with genuinely helpful AI. The divergence is interesting: one company building a product for mass consumer adoption and needing to monetise at scale; the other focusing on enterprise revenue where ads would alienate the customers paying large contracts.

For users, the practical question is: will ChatGPT ads be subtle enough that you do not notice, or will they nudge your answers in the direction of whoever paid for them? We do not know yet. But this is the moment to pay attention, because ad-supported AI is a structurally different product from subscription AI — and most people will not read the fine print.

5 Stories That Hit Closer to Home (Bangladesh)

1. Bangladesh AI Olympiad held its Dhaka regional round

On Saturday, May 2, Bangladesh University of Business and Technology (BUBT) hosted the Dhaka regional round of the 3rd Bangladesh AI Olympiad 2026, drawing hundreds of young technology enthusiasts from across the capital. The Olympiad is one of the most concrete pipelines the country has for turning curious school and university students into people who can actually build AI, not just consume it.

This matters because Bangladesh’s AI workforce gap is real. We have a large, young, tech-interested population. We have a booming freelance economy. But the number of people who can build, fine-tune, or audit AI systems domestically is still tiny compared to countries like India, China, or even Vietnam. Every student who competes in events like this — even the ones who do not win — gets something irreplaceable: the experience of seeing AI as something they can work on, not just work around.

If you have a nephew or niece who is curious about computers, look up the next round. These competitions are genuinely open.

2. AIUB opened a petaflop AI supercomputing lab — focused on Bangla

On Sunday, May 3, American International University-Bangladesh (AIUB) inaugurated an AI Supercomputing Lab at its permanent campus. The lab is not a marketing exercise. It is equipped with NVIDIA DGX Spark systems delivering up to 1 petaflop of AI performance — compact supercomputers running on NVIDIA’s Blackwell architecture with 128 GB of unified memory capable of running AI models with 200 billion parameters.

What makes this more than a hardware story is AIUB’s stated mission for the lab: building “sovereign AI” tailored to the Bangla language and local context. Researchers will use the facility to fine-tune large language models on Bangla text, build healthcare and agriculture AI tools relevant to Bangladesh, and work on infrastructure problems like Dhaka traffic and national grid optimisation.

One petaflop is not frontier scale — OpenAI and Google run at thousands of times that power. But for domestic researchers who previously had to either rent cloud time in dollars or leave the country to do serious AI work, a local 1-petaflop machine changes the economics completely. Expect the first papers from this lab by early 2027.

3. Dhaka Police is deploying 11,000 AI-powered CCTV cameras

This one requires eyes wide open. Dhaka Metropolitan Police announced this week a plan to install 11,000 AI-powered CCTV cameras across the city as part of what is being called a smart policing initiative. The cameras will use AI for real-time surveillance and pattern detection.

On the surface, this sounds like a crime-reduction tool, and in some respects it may be. Dhaka is a city of over 20 million people with genuine public safety challenges. AI-assisted surveillance has helped reduce crime in cities from London to Singapore.

But the detail that needs scrutiny is oversight. Who controls the footage? Under what legal framework can it be used? What happens if the AI misidentifies someone? Bangladesh’s draft National AI Policy 2026-2030 specifically bans social-scoring surveillance — but AI-powered mass CCTV can slide into that territory quietly and incrementally. The cameras are being installed now. The oversight framework is still being written. That sequencing is backwards, and citizens and civil society groups should be asking hard questions before, not after, the infrastructure is live.

4. Dhaka court jails ten for deepfake fraud against a religious figure

On May 5, a Dhaka court placed two men on remand and denied bail to eight others in a case that shows exactly how quickly AI-enabled fraud has escalated in Bangladesh. The group had collected images and voice recordings of popular Islamic speaker Mizanur Rahman Azhari from social media and used deepfake technology to create fabricated videos of him promoting an aphrodisiac product called “Eurocin.” The fake videos circulated across more than 24 fake Facebook pages and websites, and customers — believing they were seeing a genuine endorsement — bought the product.

This is a crime on multiple levels: fraud against the customers, defamation of the speaker, and a demonstration of how cheap and accessible deepfake tools have become. What used to require a film studio and months of work now takes a smartphone and a free app.

The good news here is that the justice system moved. Arrests were made, a court hearing happened, and remand was granted within weeks of the alleged fraud. That is faster and more decisive than many comparable cases. The bad news is that for every fraud ring that gets caught, ten others are probably running variations of the same scheme right now. The tools are too easy to access and the detection challenge is hard.

5. The “Hour of AI” campaign passes its midpoint — ask your child’s school if they noticed

If you have a child in grades 6 to 9, this is a nudge: the nationwide “Hour of AI in Bangladesh-2026” campaign launched by the Directorate of Secondary and Higher Education on April 27 has now passed its midpoint. The campaign runs until May 26 and is designed to give every secondary school student a basic introduction to coding and how AI works, supported by Code.org and the a2i programme.

The honest answer is that rollout across thousands of schools is uneven. Some schools have enthusiastic teachers who ran the session in week one. Others have not started. The lesson, in many cases, is a 45-minute introduction that will not turn anyone into an AI engineer overnight.

But that is not really the point. The point is that a generation of Bangladeshi teenagers is being introduced, officially and at scale, to the idea that AI is something you can understand and potentially build, not just something that happens to you. In ten years, when some of those students are starting companies or writing policy, that framing will matter. Ask your school. If the session has not happened yet, ask when it will.

Wrapping up

Ten stories in seven days. Globally: ChatGPT became more honest as its default, Anthropic overtook OpenAI in revenue for the first time, the Pentagon’s AI supplier list drew a controversial line, Meta launched a closed-source model while cutting a tenth of its workforce, and OpenAI opened the door to advertising inside ChatGPT.

At home: Bangladeshi students competed in the AI Olympiad, AIUB opened the country’s most powerful academic AI lab, Dhaka is getting 11,000 AI cameras with questions still unanswered about oversight, a Dhaka court moved swiftly against a deepfake fraud ring, and the “Hour of AI” campaign continues quietly rolling out across our secondary schools.

The world is not waiting for Bangladesh to catch up. But the infrastructure to participate — the labs, the education programmes, the courts that actually respond to AI misuse — is being built, piece by piece. That is not nothing.

Pick one story from today’s post. Share it with someone who does not follow tech. And if your chacha asks you what AI is, send him this post.

See you next week. Take care.