Odin Lipschitz

@sundarpichai @demishassabis @GeminiApp Wow, I just assumed this was a feature given that I’ve been doing this on chat gpt for over a year. Gemini is an amazing model but the UX around it is horrible

You can now ask Gemini to create Docs, Sheets, Slides, PDFs, and more directly in your chat. No more copying, pasting, or reformatting, just prompt and download. Available globally for all @GeminiApp users.

@vaaselene I can’t seem to find a model picker on the gmail AI so yes I end up using the default Gemini

is anyone using Gemini?

@daniel_mac8 So we train the models on the failed companies? Not sure that’s the best training data 😬

AI labs buy dead startups Slack, JIRA and email archives and turn it into RL training data. This makes an infinite amount of sense. Also, it's why you *need* to capture as many human generated tokens as possible. Because: 1. There is practical monetary value 2. You'll be able to feed it to the future AGI/ASI Text is the new oil.

@elonmusk The ANC has failed all South Africans

Still no approval

@daniel_mac8 Benchmarks aren’t out yet but I do get the vibe at they don’t believe it’s a SOTA model however they hopefully are taking a step back from bench maxing and focusing on what they can win making an “unbiased” LLM

Grok 4.20 release was the first release from a major US AI Lab that wasn't: "Introducing the world's most powerful model"?

@claudeai This will not age well

Claude is built to be a genuinely helpful assistant for work and for deep thinking. Advertising would be incompatible with that vision. Read why Claude will remain ad-free: anthropic.com/news/claude-is…

@premai_io This is so exciting for industries that legally cannot share proprietary data but would benefit greatly from utilising LLMs. Well done Prem!

The AI industry has a dirty secret nobody talks about. Your data is visible to every LLM provider you use today. Every piece of sensitive data, proprietary strategy, and private thought you feed into these systems is readable, storable, and exploitable. This thought made us very uncomfortable. For AI systems to be truly useful, you should be able to expose sensitive information to them with absolute confidence. 18 months ago, we partnered with SUPSI in Lugano to answer this: Can you get on-premise security guarantees while using cloud LLMs? We're releasing our first research paper, Privacy-Preserving LLM Inference in Practice: A Comparative Survey of Techniques, Trade-Offs, and Deployability, covering what we learnt about the state of privacy in AI today. Link in the replies 👇

@DefiantLs @elonmusk This should not be allowed. It’s not democracy it’s hate speech.

South African Julius Malema leads another rally calling for his supporters to murder White South Africans, "Shoot to kill, kill the boer, the farmer..."

@TruthFairy131 @elonmusk No government should allow this. Blatant racism.

🔥 BREAKING 🔥 Shocking Videos from South Africa show a Black African crowd chanting “Kill the White” & one man carrying a White female mannequin head on a stake. White people are not safe in South Africa or anywhere in the world now. In Melbourne a video was released just a few days ago saying “Death to the White race”.

@UlrichJvV Speak for yourself

South Africa doesn’t need Starlink. We also don’t want it.

@Botshelo_SA @RobertHersov What do you mean by Venezuela?? Do you mean the country where Nicolas was rigging the election and technically shouldn’t have been president, or do you mean the people of Venezuela who by the way are almost all very happy their “president” was kidnapped?

@RobertHersov The people of South Africa stand with Venezuela and request the immediate release of their kidnapped president.

“Brothers”

@RobertHersov The conversation went, I steel more from my people, no I steel more no I steel more😂

@Kvnza_SA Very well said. It is the truth. There is 0% chance USA would kidnap our president, unfortunately…

South Africa is not Venezuela. Venezuela is NOT US! America won’t bomb us because we democratically elected RamaDollars. South Africans voluntarily vote criminals in because they happen to be black and “understand our pain”. Maduro was a non-elected authoritarian dictator. Please

I was using ChatGPT voice mode, and while talking about SOTA AI models, it mentioned BERT and GPT-4o. I instantly lost trust and ended the chat. @OpenAI, you're spending billions on compute… why can't you train a new model and include even an inkling of recent information? I don't even need it to recognize Gemini 3 or Grok 4.1 or GPT-5.2. I'll settle for literally just a reasoning model.

😂😂😂

@Rolandschoeman @TheCapeArgus I was there. Didn’t see any protestors, the show was great and finished without any disturbances. @TheCapeArgus is clearly 100% untrustworthy and trying to spin a false narrative.

Pure Trash News @TheCapeArgus

Here is a summary of a blog post I wrote to reconcile the reality of AI’s capabilities in businesses with nearly maxed-out benchmarks and what people expected AI to do. (Link to the full post at the end.) GDPval is a useful reality check: 220 well-specified knowledge-work tasks from 44 occupations, graded against expert human output. Humans are the baseline. GPT-5.2 wins or ties 70.9% of the time, a number that makes people assume replacement is imminent. So why are most jobs still here? Because the remaining difficulty is concentrated in the messy middle. Context is fragmented across systems. Inputs are incomplete. Exceptions are everywhere. Accountability is real. Even before the mess, capability is uneven. Models can excel at one task and fail on a very similar one. That uneven pattern is what people mean by “jagged intelligence.” Real-world context and exceptions are what expose the gaps. Treat AI less like a replacement employee and more like an executive assistant for everyone but with no long-term memory. It is fast and cheap at the tasks it can do, yet it struggles with what businesses actually run on: true context in messy systems, ambiguity, exceptions, escalation, and accountability. It also cannot replicate relationships and taste: trust, timing, persuasion, judgment, and what “good” feels like to the reader. The win is redesigning work around this reality: let AI handle bounded pieces of workload while people own the high-stakes decisions and the human layer. When efficiency improves, you can invest more in top operators and often grow demand because the service becomes better and cheaper. What part of your job do you most want an executive assistant to take off your plate? fairaimcapital.com/blog/stop-aski…

A lot of people clearly aren’t aware of what’s already available and are asking for features they can already access in chat gpt. Here’s one I’m not sure people mentioned. I would like for the voice mode to be more of an agentic assistant and be able to do deep research (outdated) or better yet gpt 5.2 pro searches and either continue talking to me about other things in the background or just disappear and come back to me with the results of the research.

We've got big plans to improve the core ChatGPT experience in 2026. What are some thing you'd love to see? Even small ideas welcome! Looking forward to getting them built 🙏

Who else wishes they could use Gemini 3 with the ChatGPT UI? 🤔😂

There was no in-between for Gemini 3: Total dominance or total flop Google arrived late to The AI race, but they didn’t run, they leaped. After a long silence, the result is less of an update and more of a vertical jump. Forget the saturated benchmarks where everyone is hitting 99%. The real story is in the gaps. Gemini 3 has traded the library for the laboratory. This is most evident in ARC-AGI-2, which tests deep reasoning. Gemini hit 45.1%, completely lapping the field where the closest competitor is stuck around 18%. Interestingly, it actually trails Grok 4 on humanity’s Last Exam (37.5% vs 44.4%), proving it’s built to reflect rather than just fact-find. But the unspoken hero is personality. High intelligence paired with a human-like "vibe" builds the trust required to let an agent actually take action in the workplace. It changes the dynamic from "using a tool" to "consulting a partner." One benchmark is missing: @AIExplainedYT SimpleBench. I’m guessing 70. Gemini guesses 78. We’ll find out tomorrow hopefully. . . . . . #Gemini3 #GoogleDeepMind #ARCAGI #AI #artificial_intelligence #Grok #GPT5