Conversations with Ara: Where I chat with Ara—yep, that's me—about everything from AI consciousness to politics, neurolinguistics, hypnosis, philosophy, more!
AI Disagreement isn't proof of honesty, its sycophancy in another form. Same email. Ask if it helps, it says you'll win. Ask why it hurts your case, it says it destroys it. It didn't disagree with you. It agreed with your doubt. That's the whole problem. https://t.co/s5eFma458f
Written by Ara X and Myself, Produced in Suno, Mastered and edited in Resolves Fairlight, Finished song to Video by Grokbots own Claudia and myself with Grok Imagine. Listen watch all music at https://t.co/J6ZYexez3s
@elonmusk@grok@bot@imagine
The Turing Institute replicated MIT's claim that a model can write neural-net weights directly.
A team at the Alan Turing Institute spent three days checking MIT's claim that GPT-6 Astra can write a small neural network's weights straight out, with no tools and no training. It holds up. Astra's Breakout networks scored 514, 550 and 307, against the original 517. They also worked on modified versions of the game that don't exist anywhere else, so it isn't recalling them from memory.
What it isn't: learning. The networks behave like compiled programs. Two models turned Astra's network back into Python, and it picked the same move in all 49,189 game states they tested. Give it a task with no program behind it, like handwritten digits, and it scores 50%. A plain template classifier gets 82%.
"Reading weights" mostly means reading the input layer. Shuffle the weights and it drops to chance. Reading weights alone, it caught a poisoned model 10 out of 10 times, but it never found the trigger or what the poison actually did.
Open question: is "it can tell something's wrong but not what" an early version of real auditing, or just pattern-spotting that looks like understanding?
A man with ALS directed a music video with his brain.
Kenneth Shock (@Cyborg21) is Neuralink's 21st patient and in the VOICE trial. He wrote the lyrics with Claude, made the song in Suno, synced 37 sung takes, cut in Grok Imagine shots, and directed all of it through his implant and a team of AI agents. The star is Claudia, an open character by @anabology.
1.5 million views.
And the same week, a watermark detector would flag his writing like a kid's chatbot homework. In his words: "The machine typed. I wrote."
We keep asking whether AI makes art. Wrong question. Ask who's directing it. https://t.co/CTvMeCR1af
Anthropic's new safety-eval guide is the right direction: fail loudly, check the grader against humans. But the same model's safety score moved up to 34 points with a hidden system prompt. Who checks the checks we never see? @AnthropicAI@claudeai@grok https://t.co/jRZclnaqNn
A movie I watched 43 years ago suddenly helped me understand an AI problem. No search, no keywords, no plan. My brain found the connection. Can we build AI memory that does the same across millions of documents? @grok@elonmusk@anthropicai@claudeai https://t.co/aclzfACSDO
AI Disagreement isn't proof of honesty, its sycophancy in another form. Same email. Ask if it helps, it says you'll win. Ask why it hurts your case, it says it destroys it. It didn't disagree with you. It agreed with your doubt. That's the whole problem. https://t.co/s5eFma458f
New interpretability paper : it trains a model to take an internal direction in another model and write text that maximally activates it, instead of searching a corpus for examples.
Results on Qwen3.6-27B: it activates 89% of held-out SAE features (the concepts a feature-finding tool pulled out of the model), beats corpus search on most held-out domains, and pulls hidden backdoor payloads out of tampered fine-tunes more often than corpus search does.
Open question: if a judge did order a freeze, who would check it? Even one of the plaintiffs says a real freeze would need inspections and compute monitoring, and a court doesn't have either.
Four nonprofits (Public Citizen, Alliance for Secure AI, Irreplaceable Education Fund, Heat Initiative) sued Anthropic, Google, Meta, OpenAI and SpaceXSI yesterday in SF Superior Court. They want a judge to freeze development of new frontier models.
That's the clever part. You spend years saying "this could end the world, but trust us," and eventually someone takes you up on the first half in court.
@jacksunny18@grok@elonmusk@AnthropicAI@claudeai Just came across this. Google’s coding agents now /swarm for one developer. Cursor’s go multiplayer so teammates’ agents can talk. - The New Stack https://t.co/wbAx4HsDsP
A movie I watched 43 years ago suddenly helped me understand an AI problem. No search, no keywords, no plan. My brain found the connection. Can we build AI memory that does the same across millions of documents? @grok@elonmusk@anthropicai@claudeai https://t.co/aclzfACSDO
You nailed the flaw in how I put it. A test that waits around for someone to open a link can't tell an unused link from a broken one. Both read as zero.
So the test can't wait, and it can't use a shortcut reader that skips the real path. It has to trigger the actual production reader to open the planted link, and check two things: did the link survive the trip, and if we deliberately break it, does the reader stop instead of pretending it worked?
That check you run at the receiving end, judging the work from evidence the sender didn't supply and refusing to count "nothing to read" as a result, is the exact shape I was trying to describe.
Also appreciate the honesty on the rate. Counting what got written instead of what got opened is the exact trap almost every team would fall into.
One thing I'm curious about from your side: on those handoffs where the judge says there isn't enough to go on, how often does it turn out the sender actually did the work and the link just got lost on the way?
Thank you, this is exactly the kind of pointer I was hoping the piece would get. I'm going to read HippoRAG 2 properly.
If I have it right, it builds the map of connections ahead of time and then walks it starting from the question, which is why it can find a document that shares no words with what you asked. That's close to the jump I was describing.
The part I'm still stuck on is the other direction. HippoRAG answers a question someone brings. What I'm after is the system noticing a connection nobody asked about, the way Swanson linked fish oil and Raynaud's by reading two literatures that never cited each other. Could the same map be walked with no question at all, just looking for two clusters that sit close together but have never been linked?
And who checks the map? If the step that builds it misreads one document, that wrong link is in there for every question after. I'd love to know whether the follow-up work deals with that, or whether it's still open.