Elon Musk: OpenAI Betrayal, His Future at Tesla, and the Next Big Thing - Grokipedia

The episode opens with a running joke rather than news. One participant explains that they get "a little upset" each week and shout "Discracia". They say they bought the domain disgratia.com purely for their own amusement and announce a new recurring segment, "Disgratia Corner", introduced with a montage of movie-style insults and threats.

Another participant first protests that they missed the memo, then picks a target: Jason Calacanis (rendered "Kalkanis" in the transcript) and Pete Buttigieg. The complaint is that in the first thirty seconds of an interview the two "compared virtue signaling points" about their time at Amnesty International. The participant calls this work "affecting zero change" and says it earns a badge used "to hold over other people."

The first participant then calls a Sydney Sweeney dress that was trending on social media "too much". Others push back, one calling it "elegant", and the group jokes about why the algorithm keeps showing them the dress. That joke leads into the first real topic of the conversation.

The participants trade complaints about their feeds. One says they kept getting the dress all day, and a colleague, David Sacks, was reportedly shown posts about "SlutCon" for two weeks, which another participant confirms is "a real thing in San Francisco." The general observation is that if you show interest in anything on X, whether by clicking or liking, "the algorithm is on it" and shows you much more of it.

The reply explaining X's side describes two separate problems. The first was "an important bug" fixed over the weekend that kept in-network posts from being shown, so users who followed someone would not see that account's posts. The second was that the ranking system probably did not account for dwell time, meaning simply lingering on a post. Any active engagement, though, such as a favorite, a reply or other interaction, would bring "a torrent of that same thing."

It was overcorrecting. It had too much gain on it

Not every reaction was negative. One participant said the over-amplification had a useful side effect: it showed who else had a take on the same subject, which gave something like "a three sixty view" of whatever the user had engaged with. The response agreed but said the system had obviously gone too far. Another participant summed it up as being force-fed: one taste, and "three helpings."

Asked whether ranking now runs through Grok, the speaker said X is "gradually deleting the legacy Twitter heuristics." Removing them has exposed hidden dependencies: one heuristic's bug turned out to be covering for another bug. The speaker compared this to the internet meme of a complicated machine held up by a single small stick, and mentioned in passing that Amazon AWS East had recently had something similar.

One participant suggested blending the original reverse-chronological feed of friends with the new algorithmic feed. The response noted that everyone still has the Following tab, and that X plans to add an optional curated version of it. The problem it addresses is that some accounts post far more than others, which makes a purely chronological Following tab hard to use. The curated option would show "the most interesting things posted by your friends," with a further option for "little bit everything." The speaker also said the over-amplification bug had been "mostly" fixed, which implies some of the problem remained.

The "really big change" described is having Grok read everything posted to the platform. The speaker estimated about 100,000,000 posts per day, while noting this may count only English, with more beyond that. The plan is staged. Grok would first read what it judges to be the top 10,000,000 of those posts, then understand, categorize and match them to users. Once that scales reasonably well, it would expand to all 100,000,000 a day.

this is not a job humans could ever do.

Asked how much compute this would need, the speaker guessed it is "probably on the order of 50 k h 100," presumably about 50,000 H100 GPUs. They framed this as a rough estimate, not a measured figure. The same system would support semantic search: a user could ask a question in plain language and get matching text, pictures or video.

The conversation turns to the third anniversary of the Twitter takeover, which participants recall happened around Halloween. One participant offers a sympathetic summary. In their view the purchase "saved free speech" and was meant to stop the "woke mind virus" and make "truth exist in the world again." They add that folding X into xAI made it valuable as a data source. This is the questioner's framing, and the reply did not directly address it.

Instead, the reply pointed to two verification tools. One is Community Notes. The other is the Grok icon on X posts, which lets a user ask Grok to analyze and research a post and judge whether it is "the truth, the whole truth, or nothing but the truth" or needs context. The speaker claimed the recent Grok update is "very good at piercing through propaganda," but gave no measurement to support this.

One participant described a habit of replying to their own posts with metrics, tagging Grok, and asking whether the information is correct and whether better information exists. They called this a "truth seeking missile" and joked that Sacks then shows up in the replies to correct them. They suggested Grok could prepare its own community-note-style context on popular threads in advance, for example showing where a specific statistic comes from.

The response was that almost every post, except advertising, already has the Grok symbol, so analysis is one tap away, and that X does not want to "clutter the interface" with pre-generated explanations. The speaker stated the goal as making X "the best source of truth on the on the planet by far," and said they think it already is. That is an assertion; no evidence was offered for it in this exchange.

The speaker described the core of Community Notes as agreement across past disagreement. They added that all Community Notes code and data are open source, so anyone can recreate a note independently, and judged that "by and large, it's worked very well."

people who historically disagree agree that this community note is correct.

The speaker said the latest Grok version was also used to build Grokipedia. They described it as more neutral and accurate than Wikipedia, with more information than a typical Wikipedia page. Asked whether it was seeded with Wikipedia and how it was made, the first answer was simply "we used AI." A follow-up asked whether the process was fully unsupervised on synthetic data.

The fuller explanation was that this only recently became possible. The team finished training a "maximally truth seeking" version of Grok aimed at "cogent analysis": breaking an argument into its axiomatic elements and checking that the axioms are likely true, are not contradictory, and that the conclusion most likely follows from them. The speaker said training Grok on critical thinking "was quite hard."