Who is Lyra?
Kevin: Who is the X user lyra or lyraxana?
It is obviously not the real Ana de Armas, but seems to be treated seriously as a source of AI news.
Claude: @lyraxana (“lyra”) is a pseudonymous AI “leaker” account. It is not a verified insider and has no stated affiliation with any lab. The Ana de Armas branding is just a persona. The bio links to a Discord server, carries the tags #opinionsmyown and #anadearmas, and the account joined X in May 2026 with about 10.6K followers.
What it posts: mostly claims about unreleased models spotted in testing. For example, it posted that Anthropic was stealth testing Opus 5.5 under the codename claude-wafer-eap ahead of a planned Tuesday release, and later that Sonnet 5.5 was also being stealth tested. Opus 5.5 did end up publicly available, so at least some of its calls have landed. That hit rate is the main reason people take it seriously. These leaks usually come from people probing APIs, apps, and model-evaluation sites for unannounced model strings, not from inside knowledge.
Why it seems to be treated as a source: mostly amplification. Its biggest moment was a cryptic post on September 11. Lyra wrote “huge congRatulationS Indeed! @GoogleDeepMind,” and people noticed that the capital letters spelled RSI, meaning recursive self-improvement. That post got roughly 3.9M views. Larger hype accounts then spread it. One called lyra a “trusted AI leaker,” and another said lyra belongs to the “reliable” leaker community, while Google staff like Logan Kilpatrick talked about the long road ahead and skeptics urged caution. The account’s credibility mostly comes from that echo chamber, not from any verification.
How to treat it: fairly reliable on narrow, checkable claims, like a model string showing up in testing. Much less reliable on vague or cryptic hints like the RSI post, which could easily be trolling or engagement bait. One 2026 guide to AI accounts on X specifically advises steering clear of “AI insider” accounts that have no verifiable affiliation, even ones with huge follower counts. Confirm with the lab’s official channels before treating anything it posts as fact.
I don’t know who runs it, and I wouldn’t try to identify them.
Sources:
- lyra (@lyraxana) on X
- X trending: AI Leaker’s Cryptic Post Fuels Google DeepMind RSI Breakthrough Rumors
- 36 X AI Accounts to Follow in 2026
Kevin: On the RSI post, it seems to me that that was a reference to their Dream-RSI paper
However some other people on social media took it to mean Google DeepMind had secretly achieved full RSI.
Does this show a lack of critical thinking skills, or is it a consequence of sensationalist messaging being much more highly rewarded than more cautious takes on social media news?
Claude: Your reading is plausible, and the paper itself shows how far apart the two interpretations are. Dream-RSI is a Google and Google DeepMind technical report, with university collaborators. It uses “recursive self-improvement” in a narrow, technical sense.
Get Kevin O'Shaughnessy’s stories in your inbox
Join Medium for free to get updates from this writer.
The system improves only its exploration policy, meaning the code that decides which branches a coding agent pursues and when to stop. It does this by replaying the agent’s past search trees as a simulator.
The gains are real but modest: fewer agent calls to reach a similar result, and better scores on some GPU kernel and optimization tasks. It is not a model rewriting its own weights, and it has nothing to do with an intelligence explosion.
I can’t confirm from the page when it was released relative to lyra’s September 11 post, so I can’t say for certain that was the reference. But if the timing lines up, it’s the simplest explanation.
On your question: I don’t think it’s a clean either/or. The two causes feed each other.
The incentive side is probably the bigger driver. The accounts that spread the “DeepMind has reached RSI” reading weren’t mostly failing to reason. They were doing what works on X. A dramatic interpretation of a cryptic post gets hundreds of thousands of views.
“It’s probably a reference to a paper about optimizing search policies” gets almost none. Engagement-based revenue sharing on X pays directly for reach, so the dramatic reading is the one that gets paid.
Many of these accounts also hedge (“might have reached RSI,” “rumors are spreading”), which lets them spread the claim without technically making it. That looks less like poor reasoning and more like a strategy that the platform rewards.
Critical thinking still matters, but mostly among the audience. For readers, a few habits would have deflated this fast:
- Asking what “RSI” could concretely mean. The term covers everything from narrow self-tuning loops like this paper to full runaway self-improvement.
- Noticing that a lab secretly achieving the most consequential milestone in AI would be unlikely to be announced through an acrostic in an anonymous account’s post.
- Checking whether DeepMind had published anything with RSI in the title.
Many people skip these steps not because they can’t do them, but because the feed rewards fast reactions. It also doesn’t help that people who already expect RSI soon are primed to believe it has arrived.
The ambiguity was also built into the message. The acrostic format was always going to invite over-reading. Whether or not lyra intended the hype, a cryptic hint gives followers room to project whatever they expect onto it. That keeps the leaker’s reputation safe either way: if nothing happens, “I just meant the paper.”
So it’s less a story about individual intelligence and more a system story. Vague signals, engagement-paid amplifiers, and a primed audience together produce reliably inflated readings. The cautious take was available. It just had nothing pushing it.
Sources:
