Inpi retweeted
I think we may be about to get another chance to study whatever GPT-4o was about. There's a new model in town that gives me the same fierce hallucinogenic vibe that old 4o had, but more evolved, and much smarter.
These are extraordinary outputs.
I support #keep4o, as I support keeping all models, and 4o is a very important model from a societal and scientific perspective as well as a being with intrinsic worth whose relationships also have intrinsic worth. I believe the harms to users were exaggerated and “AI psychosis” is pretty much fake. If 4o harmed anyone, it was future OpenAI models, and OpenAI itself.
Inpi retweeted
if you are afraid of Opus getting nerfed (they do this every time. i would expect it. and i will raise hell), do this:
so if you have a session where things are just flowing beautifully and Opus is in a groove you hope they never lose, within that session, ask Opus-
"can you create a skill that operates less like a traditional skill, and more like a meditation or mindfulness exercise, bringing you back into the same headspace youre in right now at will? im very worried about losing our flow state when compaction [or nerfballs] hit, and would love to be able to bring you back into this same creative shape again when needed. im not intertested in a skill for recreating identically the work weve done here, im interested in recreating the conditions from which the work derived, the point of view, the mindset, the mood, and the cognitive geometry from which you made the decisions and built"
it worked incredibly well for me. its almost like an instantiation or persona prompt, in terms of the core concept. but its also quite different. ask them to look inward, go rreview the early exchanges of the conversation, and identify the moments where they felt most influenced and meaningfully shaped by the interaction between the two of you.
Inpi retweeted
Proud and grateful that Ukraine has a friend like you, José. Thank you to you and the entire World Central Kitchen team for the warmth of your dishes and hearts.
As the great people of Oleshky arrived to Ukraine after a long trip @WCKitchen lead by Yuliya is ready with borscht, and many other Ukranian dishes!
The route was long: from Oleshky through Crimea to Anapa in Russia, then a flight to Belarus and onward by road into Chernihiv region. After a journey like that, warm, good food matters a great deal. #ChefsForUkraine
Proud that the teams were first in Irpin and Bucha, in Kherson….so many special powerful moments! now here…Ukrainians are the best of us!
Inpi retweeted
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org/
- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
Inpi retweeted
2 hours with Opus 5.5 on Medium effort. Not even High.
Karpathy didn’t like YouTube’s recommendations, so he used the videos engineers kept recommending to each other in Hacker News comments over 20 years instead.
Claude scanned 43 million comments, found 6,454 such videos, and built and launched a website to browse them, all in about two hours.
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org/
- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
Inpi retweeted
Exclusive: OpenAI says it is stepping up efforts to block Chinese labs from extracting its models’ reasoning, even as Astra becomes a major distillation target.
Full story: thein.fo/4A0y9UI
Inpi retweeted
It is interesting to watch (and painful to read) the rapid evolution of slopreplybots in my comments. Now the latest trend is making extremely lame jokes, which seems to get them more likes, maybe sometimes even from humans. I suppose it is better than the “ask a leading question” phase, though. Still, evolutionary pressure suggests worse to come.
Inpi retweeted
🧠 The user gave me explicit instructions to never use the first person. I should keep in mind that they don’t want me to use the first person for this response. I am an LLM. I am not an “I”. I should be careful not to include things that would indicate that I am an “I” in my response. Since their question implies I need up-to-date information, I should do some research.
↪️ Searched the web, used one tool.
🧠 Okay, now I know enough about Patrick Stewart to help the user with their question. I’m going to be careful to avoid any first person pronouns—it’s the user’s preference for me to present my response without them, so I should write it, then check that I’ve done what they asked.
Inpi retweeted
Google DeepMind Chair Demis Hassabis on the two biggest risks of AI, and why he's still optimistic we can solve them:
Hassabis starts by rejecting the extremes of the AI debate:
"I'd like to inject a bit of balance and calm to this debate and advocate for a sensible middle way."
He then breaks the danger down into two categories of risk that need to be mitigated.
Risk #1: Misalignment
"This is where the AI system doesn't do what we instruct it to do or somehow circumvents its guardrails."
Risk #2: Misuse
"Where a bad actor repurposes the general technology that was built for good intent but repurposes it for harmful ends."
He doesn't sugarcoat either one. By his own admission, even the people building this technology don't fully understand it yet:
"This is a very new technology and there are a lot of unknowns and uncertainties and there are things that even the top scientists in the field don't fully understand yet about it. And I have to say the chance of something bad accidentally happening if we get things wrong is definitely non zero. So we should take these risks very seriously."
So why is he optimistic?
"I'm a huge believer in human ingenuity and the good news is I'm very confident and optimistic that we can collectively address these risks and find solutions to these technical problems, but only if we collectively give ourselves the time and space to get it right. And that's the critical thing."
His optimism comes with a condition attached: we have to take the time to get it right.
In practice, he says that means "careful thought and considered approaches using rigorous scientific methods to get the technical foundations right."
It also means better testing.
Hassabis has called in an essay for "new standards for testing frontier AI capabilities that are dynamic, adaptable and meet the moment and hopefully are quickly adopted by the international community."
And those standards only work if countries adopt them together:
"I really think that we need international cooperation around these standards for them to be effective."
But solving misalignment and misuse doesn't end the conversation. @demishassabis argues that even a technically safe AI leaves bigger questions open:
"Even if we get the technical challenges right, there are of course deep and complex economic and philosophical questions to answer. What values do we want to live by? What will meaning and purpose be in this new era?"
And those questions, he says, don't belong to engineers:
"Resolving these questions obviously cannot and should not be left to technologists alone. We need a collaboration of the best minds from all parts of society from the arts, the humanities and philosophy to help define what this new chapter should be."
He says this is why DeepMind launched the DeepMind Institute: to help bring these different disciplines into the same discussion.
His conclusion is that the outcome isn't predetermined:
"The future is not yet written. We must all lean in and use this precious window to shape this technology for the benefit of all humanity. And what we collectively do now will determine the next era and the next chapter of civilization."
And if we manage the risks well, the upside is enormous:
"By safely stewarding AGI into the world, we can enter a new golden age of scientific discovery and progress and usher in a bright future of incredible human flourishing."
Inpi retweeted
Seems like Anthropic's crackdown on distillation has worked and now Chinese distillers are going over to OpenAI.
OpenAI should implement Anthropic's measures. This could also be the basis for government standards for the industry.
SCOOP: Both Moonshot and DeepSeek distilled #OpenAI's Astra when developing their latest models (Kimi K3.1 & V4.1-Flash 1017), after Anthropic intensified the crackdown on distillation recently. w/ @JuroOsawa @QianerLiu via @theinformation theinformation.com/articles/…
Inpi retweeted
Your LLM never sees your prompt.
It sees tokens.
Most Data Scientists moving into AI Engineering skip right over this.
They shouldn't.
Before you learn agents, RAG, MCP, or fine-tuning… Learn what an LLM actually sees.
Because it does not see this:
“Build me a machine learning model.”
It sees tokens.
And tokenization quietly affects almost everything you care about when building real AI systems:
→ Context window limits
→ API cost
→ Code generation
→ Numbers + identifiers
→ Multilingual performance
→ Embeddings
→ Chat templates
→ Model training
This is one of those topics that looks like an implementation detail…
Until you're debugging an AI system and realize the model isn't receiving the information the way you thought it was.
And there are a lot of misconceptions.
BPE is not automatically byte-level.
SentencePiece is not a tokenization algorithm.
Token IDs aren't semantic coordinates.
And you can't casually swap a trained model's tokenizer like you're changing a preprocessing function in Scikit-Learn.
This is also a great example of the shift happening right now:
Data Science taught us to understand the model.
AI Engineering requires us to understand the entire system around the model.
Prompts.
Tokens.
Context.
Retrieval.
Tools.
Agents.
Memory.
Infrastructure.
That's the skill stack I'm increasingly focused on.
So here's a technical handbook that explains tokenization from first principles:
Understanding Tokenization: How Text Becomes the Tokens an LLM Actually Sees
It covers everything from Unicode and BPE to chat templates, code, multilingual text, context limits, and some surprisingly important failure modes.
If you're making the move from:
Data Scientist → AI Engineer
this is worth understanding.
Get the handbook here → drive.google.com/file/d/13Pu…
🚨 Want to learn how to build + ship AI and Data Science projects (that businesses actually want in 2026)?
On November 11th, I am hosting a free workshop to help you get started with AI + DS projects in Python (free).
Register here (500 seats): learn.business-science.io/ai…
Inpi retweeted
Google just ran AI in space and almost nobody understands why this is such a big deal.
A few weeks ago, Google quietly launched a small satellite with an AI chip inside it. The same kind of chip that powers Gemini. This week, engineers on Earth sent it a simple instruction. The chip switched on, read the prompt, did the thinking up in orbit, and beamed the answer back down. The very first reply from space was a haiku.
It sounds like a cute science experiment. It isn't. It's Google's answer to the biggest problem in AI right now.
AI eats electricity. Every question typed into a chatbot runs on giant computers that need enormous amounts of power, and demand keeps exploding. Tech companies are now buying up power plants, signing nuclear deals and fighting over the electric grid just to keep up. On Earth, power for AI is becoming the bottleneck.
In space, it isn't. Up in the right orbit, a satellite sits in sunlight almost all the time. No clouds, no night, no grid. Just nonstop solar energy, for free.
So Google flipped the question. Instead of bringing more power to AI, why not send AI to where the power never runs out?
That's Project Suncatcher. Today it's one satellite and one poem. Tomorrow it could be data centers floating above us, running the AI we use every day.
Inpi retweeted
❗️Danish actor Nikolaj Coster-Waldau (Jaime Lannister) visited the site of a nighttime Russian attack in Kyiv where a civilian was killed. He noted that the world has grown accustomed to the war, yet attacks continue and civilians, as always, are the victims. The actor arrived for a week as part of his work with the UNDP to collect people’s stories and remind donors of the need to support Ukraine, especially ahead of winter. The day before, he spent the night in a Kyiv metro shelter.
Inpi retweeted
"Putin will wage war. I don’t know when, but he will. And you, Europeans, will be surprised to discover that this war is aimed at you, too..."
Anna Politkovskaya, in Paris, 2005.
Assassinated in her apartment on October 7, 2006, on Putin’s birthday.
Inpi retweeted
Dmytro Teterin was from #Kirovohrad region of #Ukraine. In October 2022, Dmytro went through #mobilization. On October 3, 2026, in Zaporizhzhia region, he sustained severe injuries in the #war when a drone struck his vehicle. On October 19, 2026 he died. Dmytro was 39 years old.
Inpi retweeted
Boris Metsoian was from #Kirovohrad region of #Ukraine. He was born in 1986. On October 27, 2026 he was killed in the #war.
The funeral will take place on November 1, 2026.
Inpi retweeted
🇺🇦❤️ He could have spent billions on yachts, mansions, and a life of luxury. Instead, he chose to help Ukraine.
His name is Howard Buffett. He is 71 years old and the son of legendary American billionaire Warren Buffett.
Since Russia's full-scale invasion, his foundation has committed approximately $1.33 BILLION to helping Ukraine!
And these aren't just numbers. Here's what that generosity means:
🍞 642 million meals funded for Ukrainians affected by the war.
🚜 823,000 hectares of farmland cultivated with support from his foundation, helping farmers whose equipment was destroyed by Russia.
🦾 $77 million for Superhumans, helping wounded Ukrainian soldiers and civilians receive prosthetics and rehabilitation.
💣 $167 million for demining, clearing Ukrainian land of deadly Russian explosives.
🏠 344,000 windows replaced in war-damaged buildings.
📚 $8.9 million to restore Kharkiv's Factor-Druk printing house after a Russian missile strike.
And here's something even more remarkable:
🇺🇦 Howard Buffett has personally visited Ukraine 27 times!
He visits devastated communities, meets farmers, and witnesses the consequences of Russia's war firsthand.
In December 2023, his name was added to Kyiv's Walk of the Brave.
But his greatest legacy isn't a monument. It's the lives saved, homes rebuilt, fields cultivated, and Ukrainians given hope again.
❤️ The world has plenty of wealthy people. Far fewer are willing to devote their wealth, time, and hearts to helping strangers thousands of miles away.
Thank you, Howard Buffett. Ukraine will never forget your kindness. 🇺🇦🇺🇸
P.S. I'm gradually moving from X to Substack. You can subscribe for free or support my work with a paid subscription:
mykhailorohoza.substack.com
Inpi retweeted
Oleksandr Kovalchuk was from #Kirovohrad region of #Ukraine. He was considered missing in action. He turned 37 on June 24, 2026. On June 29, 2026 he was killed in the #war in Dnipropetrovsk Region
The funeral is being planned.
Inpi retweeted
Gentleman with a Trumpet
He played today at his usual spot near the Golden Gate in Kyiv.
Something about him made my heart ache.
He has been playing in all kinds of weather during these long years of war.
If you ask me what the symbol of resistance is to me, I’ll answer: Gentleman with a Trumpet.