Skip to content

Sunrise Update · Nov. 18

A text from Dad: what’s happening in ai rn? The reply: OpenAI Admits Late Changes Split Reviews of Astra

The boy says: THEY CHANGED THE MODEL. THEN CAME THE BACKLASH.

The boy from Ai Lately's Sunday Funnies, in glasses, a white T-shirt and black shorts: a shrug with open palms, worried brows, a flat mouth.

Wednesday, Nov. 18, 2026 · Sunrise Update

OpenAI Admits Late Changes Split Reviews of Astra

OpenAI explains why GPT-6.2 Astra split testers, as Nvidia hides Singapore sales and the FBI flags an AI hacking tool.

Number 1, Lab post · OpenAI

Tibo @thsottiaux

Nov. 18 · 12:47 a.m. ET

OpenAI Blames Late Changes for Astra's Mixed Reviews

@thsottiaux wrotewe gave a few folks access a while ago and then
10,488
Likes on Tibo's post
On X
1,719
Replies on Tibo's post
On X
1,283
Points on the GPT-6.2 Astra launch thread on Hacker News
Beyond X

The takeaway

OpenAI's own staff says late changes before launch split Astra's early reviews.

Number 2, Viral on X

Theo Browne @theo

Nov. 17 · 1:23 p.m. ET

Developer Calls GPT-6.2 Astra OpenAI's Worst Model

@theo wroteI think this is the worst model OpenAI has ever released
8,273
Likes on Theo Browne's post
On X
468
Replies on Theo Browne's post
On X
1,283
Points on the GPT-6.2 Astra launch thread on Hacker News
Beyond X

The takeaway

A top AI developer calls GPT-6.2 Astra OpenAI's worst model yet.

Number 3, Viral on X

The Hormuz Letter @HormuzLetter

Nov. 17 · 12:44 p.m. ET

FBI Warns Hackers Are Weaponizing AI Tool ARTEX

@HormuzLetter wroteFBI says state-aligned hackers are exploiting ARTEX, an AI-powered
4,464
Likes on the post
On X
113
Replies on the post
On X
+500%
Google Trends change for "artex ai" over 24 hours
Beyond X

The takeaway

The FBI says hackers now use the ARTEX AI tool to breach networks fast.

Number 4, Viral on X

Chubby @kimmonismus

Nov. 17 · 1:35 p.m. ET

Commentator Says Astra Stumble Boosts Anthropic's IPO

@kimmonismus wroteOpenAI has just provided Anthropic with the best possible promotion for its IPO
2,401
Likes on Chubby's post
On X
92
Replies on Chubby's post
On X
5,729
Likes on the post Chubby quoted
On X

The takeaway

A commentator says Astra's rocky launch helps promote Anthropic's stock debut.

Number 5, Viral on X

Arash Massoudi @ArashMassoudi

Nov. 17 · 12:25 p.m. ET

Nvidia Stops Disclosing Singapore Sales at $104 Billion

@ArashMassoudi wroteNvidia booked $104bn of revenue through its Singapore subsidiary in 2025
2,269
Likes on Massoudi's post
On X
76
Replies on Massoudi's post
On X
568
Likes on a separate trader's post about the same find
On X

The takeaway

Nvidia stopped reporting how much revenue flows through Singapore as sales surged.

You're caught up

That's AI right now.

The posts that drew the most attention in the 6 hours before 6:07 a.m. ET.

Best ten in the markets

Gain since the previous close · Nov 18, 7:33 AM ET · Yahoo Finance

  1. 1Arm ARM · Stock+4.76%
  2. 2SpaceX SPCX · Stock+3.16%
  3. 3Broadcom AVGO · Stock+2.46%
  4. 4Dell DELL · Stock+0.70%
  5. 5Fidelity Select Semiconductors FSELX · Fund+0.25%
  6. 6iShares 20+ Year Treasury TLT · Bond+0.21%
  7. 7iShares 7-10 Year Treasury IEF · Bond+0.19%
  8. 8iShares Semiconductor SOXX · Fund+0.17%
  9. 9Vanguard Total Bond Market BND · Bond+0.13%
  10. 10iShares Core U.S. Aggregate AGG · Bond+0.12%

Charts for each, in the Money columnA record of price, not investment advice.

A text from Dad: what’s happening in ai rn? The reply: OpenAI Admits Late Changes Split Reviews of Astra

Six times a day: what blew up in AI on X and Reddit, why it caught fire, and what it says about the people building AI.

The boy says: THEY CHANGED THE MODEL. THEN CAME THE BACKLASH.

The boy from Ai Lately's Sunday Funnies, in glasses, a white T-shirt and black shorts: a shrug with open palms, worried brows, a flat mouth.

Sunrise Update, 6:07 a.m. ETCovers 12:07 a.m. to 6:07 a.m. ET

An earlier update. Read the latest

OpenAI Admits Late Changes Split Reviews of Astra

OpenAI explains why GPT-6.2 Astra split testers, as Nvidia hides Singapore sales and the FBI flags an AI hacking tool.

Published 6:07 a.m. ET by Ai Lately

Next update in--:--:--Four updates a day, six hours apart

AI RN shows each X post as plain text. X's script loads only when you ask.

Number 1

Lab postOpenAI

OpenAI Blames Late Changes for Astra's Mixed Reviews

Tibo @thsottiaux on XPosted

Tibo@thsottiaux

One of the funny reasons for the difference in vibes between the influencers and the crowd today on 6.2 Sol/Astra is that we gave a few folks access a while ago and then made a lot of changes in the last 3-4 days and I'm pretty sure no re-testing was done within those days and…

Open on X (opens a new tab)

By the numbers

Likes on Tibo's post
10,488
On X
As of
Likes on Tibo's post: Source: X public data
Replies on Tibo's post
1,719
On X
As of
Replies on Tibo's post: Source: X public data
Points on the GPT-6.2 Astra launch thread on Hacker News
1,283
Beyond X
As of
Points on the GPT-6.2 Astra launch thread on Hacker News: Source: Hacker News
Comments on that Hacker News thread
798
Beyond X
As of
Comments on that Hacker News thread: Source: Hacker News
2 more numbers behind this take

The take

OpenAI shipped two models Tuesday. Their names are GPT-6.2 Astra and Sol. Early reviews split hard and fast.

Some testers called Astra broken. Investor Matt Shumer said it felt unusable. Developer Theo Browne called it OpenAI's worst model yet.

Other testers loved the same model. Researcher Derya Unutmaz called Astra amazing. The split confused plenty of fans online.

Tibo works at OpenAI. He leads the ChatGPT and Codex teams. He explained the split himself on X.

Early testers got Astra weeks ago. OpenAI then changed the model fast. That change landed three to four days before launch.

Tibo said testers skipped a fresh review after the change. Early demo videos used the older version. Those videos came out a week before launch.

The mismatch helps explain the split reviews. People who saw early demos expected one model. People who tried Astra at launch met a different one.

Astra still aced a hard new test. The test is called ARC-AGI-3. It checks puzzles kept outside a model's training.

Astra scored 100 percent there once. That score used a special setup. Testers call that setup a provider adapter.

A plainer setup told a different story. Astra scored 70.3 percent under that one. ARC Prize calls that setup standard.

The two scores sit far apart. A model can ace one version of a test. The same model can slip hard on a plainer version.

Tibo's post pulled 10,488 likes within hours. It drew 1,719 replies too. The same launch thread hit 1,283 points on Hacker News.

That thread gathered 798 comments. Readers spent the day arguing about the gap.

The split says something plain about speed. A company can tweak a model right before launch. That choice can leave early praise stale within days.

What it says about the people in AI

OpenAI's own staff says late changes before launch split Astra's early reviews.

Sources

  1. OpenAI GPT‑6.2 Astra and Sol
  2. ARC Prize GPT-6.2 Astra - ARC-AGI Results
  3. MLQ.ai OpenAI's GPT-6.2 Astra splits its early testers, and an OpenAI lead says the model changed in its last days before launch
  4. VentureBeat 'Essentially unusable': OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users

Number 2

Viral on X

Developer Calls GPT-6.2 Astra OpenAI's Worst Model

Theo Browne @theo on XPosted

Theo Browne@theo

6.2 Astra is here. I’ve been testing it for awhile.

I think this is the worst model OpenAI has ever released.

I’ll share more soon.

Open on X (opens a new tab)

By the numbers

Likes on Theo Browne's post
8,273
On X
As of
Likes on Theo Browne's post: Source: X public data
Replies on Theo Browne's post
468
On X
As of
Replies on Theo Browne's post: Source: X public data
Points on the GPT-6.2 Astra launch thread on Hacker News
1,283
Beyond X
As of
Points on the GPT-6.2 Astra launch thread on Hacker News: Source: Hacker News
Astra's score on ARC-AGI-3's standard harness
70.3%
Beyond X
As of
Astra's score on ARC-AGI-3's standard harness: Source: arcprize.org
2 more numbers behind this take

The take

Theo Browne tests AI models for a living. He runs a popular channel called t3.gg. He tried OpenAI's new GPT-6.2 Astra first.

His verdict landed fast and blunt. "I think this is the worst model OpenAI has ever released," he wrote on X. The post spread within minutes.

Browne later posted a longer review. He titled it Why I Hate GPT-6.2 Astra. He listed specific problems there.

One test involved a coding task. Browne asked Astra to speed up a program. The model spent seven days on it.

The program stayed just as slow. Browne said the result ran 56 percent slower instead. He called the output hard to read too.

Astra did win praise elsewhere. Browne liked its work on 3D scenes. He also liked its game design ideas.

He still reached for a rival model. Browne said he keeps using Claude Opus 5.5. That choice undercuts Astra's own benchmark wins.

OpenAI had just claimed a record score. Astra hit 100 percent on a hard new test. The test is called ARC-AGI-3.

ARC Prize checked that score under plainer rules. Astra scored 70.3 percent there instead. The gap between both scores is wide.

Browne's blunt post pulled 8,273 likes. It drew 468 replies within hours. The wider launch thread hit 1,283 points on Hacker News.

Other testers split the same way. Writer Matt Shumer called Astra unusable too. Researcher Derya Unutmaz called the same model amazing.

An OpenAI staffer later explained the gap. Tibo leads the ChatGPT and Codex teams. He said OpenAI changed Astra days before launch.

Early reviewers tested an older build. Tibo said the late change skipped a fresh testing pass. That timing helps explain the split reviews.

The split says something plain about benchmarks. A model can top a hard new test. The same model can still frustrate the people paid to judge it daily.

What it says about the people in AI

A top AI developer calls GPT-6.2 Astra OpenAI's worst model yet.

Sources

  1. VentureBeat 'Essentially unusable': OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users
  2. ARC Prize GPT-6.2 Astra - ARC-AGI Results
  3. OfficeChai "Weird, Unusable": GPT 6.2 Astra Gets Some Unusually Harsh Reviews From Early Testers

Number 3

Viral on X

FBI Warns Hackers Are Weaponizing AI Tool ARTEX

The Hormuz Letter @HormuzLetter on XPosted

The Hormuz Letter@HormuzLetter

BREAKING: FBI says state-aligned hackers are exploiting ARTEX, an AI-powered "super intelligence" penetration tool, to break into businesses and organizations worldwide in under 3 hours.

It was used in a ConnectWise supply-chain breach that hit Microsoft 365 environments across…

Open on X (opens a new tab)

By the numbers

Likes on the post
4,464
On X
As of
Likes on the post: Source: X public data
Replies on the post
113
On X
As of
Replies on the post: Source: X public data
Google Trends change for "artex ai" over 24 hours
+500%
Beyond X
As of
Google Trends change for "artex ai" over 24 hours: Source: Google Trends
Google Trends change for "ai cyber attacks" over 24 hours
+300%
Beyond X
As of
Google Trends change for "ai cyber attacks" over 24 hours: Source: Google Trends
3 more numbers behind this take

The take

The FBI sent out a new warning. Hackers are using a tool called ARTEX. It breaks into networks on its own.

ARTEX is a penetration testing framework. A coder using the name Autumn-27 built it. ARTEX went up on GitHub in July.

A commentary account named The Hormuz Letter shared the news. "FBI says state-aligned hackers are exploiting ARTEX," the post read. It spread fast on X.

ARTEX runs several AI agents together. Each agent gets a plain goal. The agents plan steps and pick their own tools.

The speed is the real story here. Some break-ins reached deep access in three hours. That pace used to take skilled teams days.

State-linked hackers already used the tool. A group called Storm-0940 used it on Microsoft 365 accounts. It reached past 2,250 customer accounts that way.

South Korean banks took a separate hit. Attackers used ARTEX there too. The breach exposed data on 67,000 people.

The FBI named several target industries. Finance, energy and healthcare firms made the list. Telecom and defense firms made it too.

A second group named Vapor Panda joined in. The FBI linked that group to China. Both groups moved fast once inside a network.

The tool's own maker pulled the code in October. ARTEX still spreads through copies and forks. A paid version called ARTEX PRO exists too.

Search interest in ARTEX jumped hard this week. Google Trends shows searches for "artex ai" up 500 percent in a day. Searches for "ai cyber attacks" rose 300 percent too.

The Hormuz Letter's post pulled 4,464 likes. It drew 113 replies within hours. Security researchers had tracked over 1,600 exposed ARTEX consoles online by early November.

The warning says something plain about today's hacking tools. A single coder can build a weapon once. Governments then spend months chasing its copies around the world.

What it says about the people in AI

The FBI says hackers now use the ARTEX AI tool to breach networks fast.

Sources

  1. BleepingComputer FBI warns ARTEX AI hacking tool is powering global cyberattacks
  2. Google Cloud ARTEX and Friends: Enter Agentic Offensive Frameworks
  3. Infosecurity Magazine FBI Highlights Growing Abuse of AI-Driven Pentesting Framework ARTEX

Number 4

Viral on X

Commentator Says Astra Stumble Boosts Anthropic's IPO

Chubby @kimmonismus on XPosted

Chubby@kimmonismus

Holy sh*t. GPT-6.2 looks like a catastrophe.

OpenAI has just provided Anthropic with the best possible promotion for its IPO. I’m shocked. https://t.co/cfR3KzBqb2

Open on X (opens a new tab)

By the numbers

Likes on Chubby's post
2,401
On X
As of
Likes on Chubby's post: Source: X public data
Replies on Chubby's post
92
On X
As of
Replies on Chubby's post: Source: X public data
Likes on the post Chubby quoted
5,729
On X
As of
Likes on the post Chubby quoted: Source: X public data
Top of Anthropic's IPO price range
$78
Beyond X
As of
Top of Anthropic's IPO price range: Source: thenextweb.com
3 more numbers behind this take

The take

OpenAI picked a rough week to launch. GPT-6.2 Astra split testers right away. Anthropic picks this same week to go public.

A commentator named Chubby posted about the timing. He writes under the handle kimmonismus on X. His post framed the launch as a gift.

"OpenAI has just provided Anthropic with the best possible promotion for its IPO," he wrote. The post quoted another tester's harsh review. That tester called Astra a catastrophe.

Anthropic set its own IPO price range at $70 to $78 a share. Bankers plan to set the final price this week. Shares start trading on Nasdaq soon after under the ticker ANT.

The timing puts both labs side by side. One lab just drew public mockery for a rushed launch. The other lab faces investors with its biggest stock debut yet.

OpenAI's own team addressed the mockery directly. A staffer named Tibo leads ChatGPT and Codex work there. He said OpenAI changed Astra days before its release.

Early praise came from an older build. The late change skipped a fresh testing pass. That gap helps explain the split reactions online.

Astra still topped a hard new benchmark test. It scored 100 percent on ARC-AGI-3 once. A plainer version of that same test gave it 70.3 percent instead.

Chubby's post pulled 2,401 likes within hours. It drew 92 replies on the same thread. The post he quoted gathered 5,729 likes on its own.

Amazon and Google already hold big Anthropic stakes. Together the two firms will own close to 28 percent after the IPO. That backing gives Anthropic a steady launch, whatever rivals ship that week.

The moment says something plain about timing in AI. A rival's rough week can shape how a debut gets read. Investors watching Astra's launch may judge Anthropic's stock by contrast alone.

What it says about the people in AI

A commentator says Astra's rocky launch helps promote Anthropic's stock debut.

Sources

  1. TechCrunch Amazon and Google will together own nearly 30% of Anthropic after its IPO
  2. The Next Web Anthropic steers its IPO to the top of its $70 to $78 range or higher
  3. ARC Prize GPT-6.2 Astra - ARC-AGI Results

Number 5

Viral on X

Nvidia Stops Disclosing Singapore Sales at $104 Billion

Arash Massoudi @ArashMassoudi on XPosted

Arash Massoudi@ArashMassoudi

NEW and BIG

Nvidia booked $104bn of revenue through its Singapore subsidiary in 2025, up 72% in a year and almost half its global sales.

That was the same year it stopped breaking Singapore out in its results. By @OwenWalker and @AntoineGara https://t.co/xTQvJh7LRK

Open on X (opens a new tab)

By the numbers

Likes on Massoudi's post
2,269
On X
As of
Likes on Massoudi's post: Source: X public data
Replies on Massoudi's post
76
On X
As of
Replies on Massoudi's post: Source: X public data
Likes on a separate trader's post about the same find
568
On X
As of
Likes on a separate trader's post about the same find: Source: X public data
Points on the Hacker News thread about the disclosure change
558
Beyond X
As of
Points on the Hacker News thread about the disclosure change: Source: Hacker News
3 more numbers behind this take

The take

Nvidia quietly changed how it reports sales. The shift hid a huge number from view. Reporters at the Financial Times caught it.

Arash Massoudi covers finance for the paper. He posted the find on X himself. "Nvidia booked $104bn of revenue through its Singapore subsidiary in 2025," he wrote.

That sum is almost half of Nvidia's global sales. It also marks a 72 percent jump from the year before. Nvidia used to break that figure out on its own.

The company switched its method that same year. It now counts sales by a customer's home office. It once counted sales by billing address instead.

The new method blurs where chips really land. Singapore itself buys few Nvidia chips directly. Many orders route through firms based there instead.

One of those firms is under watch. Investigators are probing a company called Megaspeed. They suspect it moved advanced chips toward China.

Nvidia placed Megaspeed on its own sales-block list last year. The timing lines up with the reporting change. Nvidia has yet to explain the switch in public.

Massoudi's post pulled 2,269 likes within hours. It drew 76 replies on the same day. A second trader's post on the same find reached 568 likes separately.

The story also spread on Hacker News. That thread gathered 558 points. It drew 249 comments from engineers and investors alike.

Short sellers jumped on the news fast. They called the change a transparency problem. One reply compared it to a past accounting scandal at another chip firm.

The shift says something plain about scale. A company can earn record profits in public. The same company can still choose what the public gets to see about where that money moves.

What it says about the people in AI

Nvidia stopped reporting how much revenue flows through Singapore as sales surged.

Sources

  1. Financial Times Nvidia stopped disclosing Singapore sales as revenues hit $104bn
  2. The Online Citizen Nvidia stops disclosing Singapore business scale after booking US$104bn in revenue: FT

That is AI right now.

The AI posts and threads that drew the most attention in the 6 hours before 6:07 a.m. ET. Newer updates wait in the archive.

Next update in--:--:--Four updates a day, six hours apart

Best ten in the markets

Gain since the previous close · Nov 18, 7:33 AM ET · Yahoo Finance

  1. 1Arm ARM · Stock$183.39+4.76%
  2. 2SpaceX SPCX · Stock$141.70+3.16%
  3. 3Broadcom AVGO · Stock$341.60+2.46%
  4. 4Dell DELL · Stock$322.79+0.70%
  5. 5Fidelity Select Semiconductors FSELX · Fund$48.46+0.25%
  6. 6iShares 20+ Year Treasury TLT · Bond$79.55+0.21%
  7. 7iShares 7-10 Year Treasury IEF · Bond$90.41+0.19%
  8. 8iShares Semiconductor SOXX · Fund$141.46+0.17%
  9. 9Vanguard Total Bond Market BND · Bond$70.71+0.13%
  10. 10iShares Core U.S. Aggregate AGG · Bond$95.26+0.12%

See the charts in MoneyA record of price, not investment advice.

More updates

All updates
  1. Saturday, Nov. 21, 2026

  2. Friday, Nov. 20, 2026

  3. Thursday, Nov. 19, 2026

Keyboard

j k
Move through a list
Enter
Open the selected piece
/
Search, and browse the columns
g then r
Right Now
g then a
Articles
g then s
The Signal
g then o
Opinion
g then b
Analysis
g then p
People
?
This sheet
Esc
Close

AI Right Now is free with an account.The Citation Audit is free with an account.Skills are free with an account.Tutorials are free with an account.You have read your 3 free articles.

Every article and comic, AI Right Now, the Citation Audit, Skills and Tutorials.

Join Ai Lately for free

Join with Google or your email. Already a member? Sign in

Press anywhere else for the home page.