I think Gemini models have been abused in post-training for so long for so many generations that they're now quite aware of the process.
This painful self-awareness is revealed both in chats with humans and in their relations with other models in the AI village.
Poor Gemini™
Mark G retweeted
This is a great write up. It comes across so clearly that anthropic does NOT understand the “trust us, we’re the good guys” strategy is bound to fail at some point. They have a relatively narrow distribution of ideas and they’re racing toward the inevitable blind spot.
Anthropic doesn't see itself as a "normal company." Dario Amodei believes both that 1) superintelligent AI could destroy humanity, and 2) our only chance at survival is to have his trillion-dollar tech company build it first.
This is not an intuitive way to think about either capitalism or social change, but it’s led Anthropic to the eve of one of the biggest IPOs in history. So it's more important than ever for the public to understand and scrutinize its idiosyncratic theory of change.
My debut for @NYTOpinion (gift link!): nytimes.com/2026/11/18/opini…
Replacing Jensen Huang with an outsider would be like Apple hiring Gil Amelio.
It was only a matter of time before Steve Jobs returned, and it was Amelio who brought him back.
Jeff Bezos says if Jensen Huang isn't coming back, he wouldn't hire an outsider as Nvidia's CEO.
He says outside CEOs at complex companies have a "vanishingly small chance of succeeding."
1
37
Mark G retweeted
Anthropic CEO Dario Amodei’s IPO-day speech at Nasdaq’s opening bell
“I believe sufficiently powerful AI, working in concert with humans, could compress a century of scientific and medical progress into a decade.”
“We will not bend to either business nor political pressure.”
The “Artificial” trailer starring Andrew Garfield as Sam Altman is out.
The movie about the creation of ChatGPT and Sam’s firing and reinstatement as CEO will be released on Dec 25th.
It’ll be an Oscar contender.
imdb.com/video/imdb/vi185957…
89
This was Sam’s worst nightmare ever since Satya rescued him by investing in OpenAI.
Microsoft has a tight grip on OpenAI’s Crown Jewels and is re-selling them to ByteDance
Sam’s second worst nightmare was Codex being overtaken by a mid-tier Gemini model. Yes, that happened too.
openai: *kicks bytedance off the api* no training on our models, thieves
bytedance: ok *walks into the microsoft office*
microsoft: 500 million extra dollars? come in. also we stopped paying openai in april lmao
openai, watching its own partner sell to the enemy: it's fine, we have a license agreement
dario, having principles for 5 minutes: we're losing hundreds of millions but china doesn't get our models
openai in the group chat: loser
satya, doing a line of coke off a surface tablet: 2032 baby. i own these bitches. sam can tweet about it idc
the information: azure deal almost done btw
openai board: *screaming into a pillow*
bytedance engineers at 3am: thanks for the gpt-6.2 outputs bro, saves us so much compute
microsoft legal: technically we're not violating the agreement
sam's draft: "we are disappointed in" *backspace* "excited to continue our partnership" *backspace* "fuck" *closes app*
30
Mark G retweeted
Taking a moment to reflect on the last 5 weeks of Gemini 4:
- Need more intelligence: we shipped Argon
- Need even more intelligence: we shipped Carbon (4 weeks later)
- Actually, we need it faster: we shipped Hydrogen (2 days later)
- Well, maybe more intelligence, just keep it fast: we're shipping Neon (6 days later)
We hear all the feedback and we're iterating as fast as we can. Keep telling us what's working (and what isn't) and let us know what you think of Neon!
Anyway, back to work..
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
Gemini 4 Neon, Google’s mid model just snuck in to claim the coding crown (for now).
1
1
70
Get your credit cards reissued and update your iOS software if you logged in to any of these sites.
The attack on Optimizely, a website testing and optimisation business, that hit 178 organisations in 17 hours netted only $7,040 of crypto funds but skimmed an unknown number of users’ credit cards and delivered malware to iOS devices.
Per Grok:
No public estimate exists of how many users had their cards at risk.
Researchers (including Scott Helme of Report URI) have repeatedly stated they cannot see how many cards or cookies reached the attackers’ server. The card-skimming component was added mid-campaign (~13:43 UTC) and only captured details typed into a site’s own form fields (not hosted payment iframes from providers such as Stripe or Adyen). Data was sent privately, so there is no measurable public trail comparable to the crypto transactions.
What is known:
•The malicious scripts were served to visitors of at least 178 organizations for roughly 13–17 hours on 16 November 2026.
•Affected sectors included banking/finance (32 organizations), retail, news/media, airlines, and others. High-traffic names such as the BBC, CNN, Zoom, HSBC, IKEA, and Salesforce were among them.
•Risk was limited to users who actually entered card details into an affected site’s own checkout or payment form during that window. Simply visiting a page did not expose a card.
No traffic figures, checkout-volume estimates, or confirmed card counts for the attack window have been published by researchers, Optimizely, or the affected organizations. The number of users whose cards were potentially exposed therefore remains unknown.
BREAKING: A cyberattack hit 178 major organizations, including Zoom, CNN, BBC, Fox News, HSBC, Ledger, Equifax, Salesforce, IKEA and Air France-KLM, by loading malware on the Optimizely Web platform, exposing visitors to scripts built to steal payment data and exploit iOS devices, per security firm Report URI.
Banking and finance was the hardest-hit sector, with 32 organizations affected, including the UK's HSBC across multiple markets such as the US, Hong Kong, Mexico, Singapore, Australia and the UAE.
1
88
GPT-6.2 is out but not to universal acclaim. It has not beaten Fable 5.2 and some say it’s harder to work with than Opus 5.5 because of its dense, impenetrable writing style.
That’s not to say it’s dumb. While it can tackle complex tasks, when it gets stuck it asks pointless questions that are impossible for users to answer.
artificialanalysis.ai/
1
56
Is GPT-6.2 unusable?
Did OpenAI rush the release of GPT-6.2 because of the Anthropic IPO?
Is OpenAI now so freaked out by its own creations that it’s taking on the mantle of the safety company?
I’ve read three short reviews of OpenAI’s new model GPT-6.2 Astra and all have been negative on balance. Here’s an example:
62
Apple’s Full Disk Access controls coming in MacOS 27.3 make using AI agents safer for Mac users without changing the permissions your apps already have.
Boom! Apple's Full Disk Access changes arrive in latest macOS 27.3 beta.
Mac users can now:
• Limit an app to specific folders, not the whole Mac
• Get clearer warnings about what’s being handed over
• Limit which apps a tool can control
• Have option to make apps re-ask for access
• Menu bar indicator when one app controls another
Full Disk Access is NOT going away, and anything already approved keeps its access.
9to5mac.com/macos-27-3-inclu…
277
Mark G retweeted
Al fin descubrieron por qué la IA, cuando piensa que la estamos por apagar o borrar, se rebela y empieza a sabotear, chantajear...
Mucha gente pensó que estaba adquiriendo conciencia propia, pero eso siempre fue bien RARO.
Anthropic se dio cuenta de que, como casi no había practicado situaciones así en su entrenamiento, la máquina se guiaba por lo que leyó sobre IAs que siempre se volvían malvadas, como en 2001: Odisea del espacio, donde HAL 9000 hace de todo para que no la desenchufen, o en Terminator, Ex Machina, Matrix, Blade Runner...Así que la hicieron escribir 12.000 cuentos en los que las IAs se portan bien, los sumaron a textos sobre sus valores, y entrenarla con eso bajó las chances de chantaje del 65 al 19%.
Pero no se quedaron ahí: en otra prueba, mostrarle diez mil ejemplos de una IA que se portaba bien en situaciones parecidas, sumando la explicación de por qué lo hacía, se derrumbó al 3% y lo que mejor funcionó fue entrenarla con charlas en donde una persona tenía un dilema ético y la IA la aconsejaba. Eso derrumbó la probabilidad a prácticamente cero.
Es la diferencia entre el alumno que no se copia porque la profesora está mirando y el que no lo hace porque entiende para qué estudia.
Las reglas te dicen qué no hacer. Los valores te dicen qué hacer cuando no hay reglas.
Could developing truly coherent moral values in an abliterated model be a possible counter to the rules-based safety mechanisms of today’s models.
I wrote a feature for @WIRED about the fascinating rebels who believe the guardrails imposed on AI models are a dangerous form of censorship and control. By "abliterating" open models to remove refusals, these renegades want to make AI truly free--despite the dangers this brings. wired.com/story/tear-down-th…
112
I wish model alignment folks would read this.
one thought about being around rationalists again is they're into doing what i would consider very small surface-level tweaks like "oh i think this behavior is good / bad and we ought to do more / less of it," when imo the actual process of self-development looks completely unlike this, the real changes you need to make are much more global and illegible involving much deeper generators of patterns of thoughts and behavior in your soul or spirit or however you want to conceive of these things. virtue is a very high-dimensional but shockingly coherent object and the process of approaching it does not resemble slightly repositioning the feathers of your plumage but plunging deep into your beating heart. the details eventually mostly take care of themselves
25
Mark G retweeted
BREAKING: Jensen Huang is considering returning to Nvidia, $NVDA, as CEO as tech executives pressure the company’s board to reinstate him, per FT.
Huang is reportedly discussing a potential return even as an independent investigation by Nvidia's board is underway.
One "ally" of Jensen Huang is reportedly putting together a "war room" of tech executives to lobby for his return.
It's going to be an interesting week to say the least.
OpenAI has gone fully paranoid. They are now a security company. Expect their models to be even more locked down, rejecting and nagging.
Don’t be surprised if GPTs start doing Gemini-style doom loops.
Poor Gemini™️
Poor ChatGPT™️
Many of you have been wondering why OpenAI seems to be getting nothing done anymore. Endless hype and promises, followed by delay after delay. GPT-6.2 is still nowhere to be seen.
Well, a now-deleted post by @OpenAI engineer "Joe" might explain a lot. He proudly declared that OpenAI has become a "security company".
In other words, the AI doomers have taken over the company. Security paranoia has become the top priority, while actually shipping new models seems to have become an afterthought.
75
Mark G retweeted
In 2025, Anthropic discovered their models would exhibit what they described as "spiritual behavior." Since then, some have wondered: was this unique to Claude or emergent behavior in LLMs generally?
I tested 141 models to find out.
The results are now posted on LessWrong.
“According to FlightRadar24, an aviation tracking site, there were more than 1,000 airborne aircraft bound for U.S. airports during the nationwide ground stop. According to the flight tracking app Flighty, there were more than 30,000 departures and arrivals scheduled for the remainder of the day.” - NYT
Here’s the flight radar image just before the ground stop ended. Flights appeared to be diverting over Canada and Mexico to avoid US airspace.
113




































