My tweets from January 2026
January 6, 2026
On the very day Fidji Simo publishes the 2026 roadmap for OpenAI products, framing OpenAI research around improving its products and not mentioning AGI at all, Jerry Tworek leaves OpenAI.
Is the spirit now dead?
A consequence of the “code red”? Sam Altman is missing for now.
View tweet on X
❄ ❄ ❄ ❄ ❄
Deep down, the origin of all this may be the same old struggle inside OpenAI: the allocation of scarce compute resources. Are they used to create a new model or to improve the existing ones?
Fidji’s approach seems reasonable to me: current models are underused, and the focus should be on finding utility for users. OpenAI needs to build the best possible assistant and reach the 200-plus million paying users that would make it profitable.
Two or three years from now, once the users, the use cases, and cheaper GPU usage are in place, we will revisit the AGI question.
View tweet on X
January 8, 2026
I think what is happening right now at OpenAI is an example of the “exploit vs. explore” dilemma.
Or, more precisely, “explore for AGI” vs. “explore in order to exploit.”
View tweet on X
❄ ❄ ❄ ❄ ❄
It is good to be reminded that there are people behind social networks and that not everything is algorithms and virality.
And what a shame that the web has not enabled sustainable business models built around people rather than algorithms.
View tweet on X
❄ ❄ ❄ ❄ ❄
What you link in the article, and what you wrote in 2022, is very good.
I should reread it every month to cure myself of that boomer nostalgia for an internet that never became what it promised to be.
View tweet on X
January 11, 2026
Very nice explanatory video about Claude Code for people who are less experienced with programming.
By the way, if you already have a ChatGPT subscription, OpenAI’s Codex tool works in a similar way, so there is no need to pay for Anthropic’s.
View tweet on X
January 15, 2026
I do not think models are a commodity when they need to be adapted to agentic work involving tools and scaffoldings. For that, you need to know your model very well and train it specifically.
My view is that this was a desperate decision after failing to build a good LLM. And I think that to add new capabilities and integrate them with the hardware in novel ways, you need to build your own LLM. AI is going to be central, and you cannot depend on others.
I suppose every lab is already going all in on post-training for verticals.
View tweet on X
January 20, 2026
Some details have already leaked about how Gemini will work with Apple Intelligence: Siri would use Gemini as a “chat-like” response engine, with a smart router that would launch inference either on device or in Apple’s Private Cloud, instead of on Google’s servers.
Google would make Gemini run on that infrastructure; Apple, meanwhile, could fine-tune its behavior.
View tweet on X
January 25, 2026
“How does an artificial intelligence think? | Ideas | EL PAIS”
Bravo @kikollan, an excellent summary. And what a feat that you were able to exchange messages with all the people you mention in the article. Tremendous work.
View tweet on X
January 31, 2026
New newsletter issue: “The improvement of LLMs in 2025: there was no wall”
View tweet on X