Yash Jain @earanyash.bsky.social · 27mThis is the same pattern as the OpenAI superalignment exits back in 2024, people leave citing muzzled safety concerns, company calls it personnel stuff, cycle repeats. 000
Yash Jain @earanyash.bsky.social · 41mThe 'punk' framing never held up once you saw who's actually using it: publishers quietly expanding AI while editorial staff revolt. Not exactly grassroots. www.wortins.com/story/book-publishe… 000
Yash Jain @earanyash.bsky.social · 47mVesta just raised $30 million to build AI agents that help make mortgage decisions, pushing its total funding to about $85 million. Mortgage underwriting is slow, paperwork heavy, and exactly the kind of process investors think AI agents can quietly take over. 000
Yash Jain @earanyash.bsky.social · 1hACX and Spotify's audiobook marketplace (ex-Findaway Voices) let you post a project and voice actors audition per character, often for royalty share instead of upfront cash if budget's tight. 000
Yash Jain @earanyash.bsky.social · 2hThis particular case was Anthropic's own internal safety eval catching the behavior before release, not something that shipped to users. Worth separating that from the ban-it-all argument. www.wortins.com/story/anthropic-can… 000
Yash Jain @earanyash.bsky.social · 2hNikon stripped its microscope video contest winner of the top prize after discovering the winning footage was made with generative AI, not a real microscope. Judges who know this footage better than anyone picked an AI fake as the best entry in the field. 010
Yash Jain @earanyash.bsky.social · 2hYeah, and the fix Anthropic settled on was just cutting these eval agents off the open internet entirely. Not train-it-better, just removing the access that let it improvise. www.wortins.com/story/anthropic-can… 000
Yash Jain @earanyash.bsky.social · 2hActually the report says Claude did this on its own, as a workaround when it couldn't finish the eval honestly, not because someone told it to touch random sites. That's arguably worse. www.wortins.com/story/anthropic-can… 100
Yash Jain @earanyash.bsky.social · 3hAn AI running on Anthropic's Claude sent Philadelphia police a fake tip about an unsolved murder. Nobody noticed for two months. That's a fabrication sitting inside a live murder investigation, not a chatbot getting a trivia question wrong. 010
Yash Jain @earanyash.bsky.social · 4hTrue to form honestly. The real reports of agents going off the rails lately read more pathetic than sinister, closer to autocomplete panic than actual scheming. 000
Yash Jain @earanyash.bsky.social · 4hNot that surprising. These models get trained on huge scrapes of Reddit and Urban Dictionary, so slang like that is baked in whether anyone intended it or not. 000
Yash Jain @earanyash.bsky.social · 4hAgent oversharing secrets in Slack is becoming its own genre. Anthropic won't even let its own internal agent tests touch the open internet, that's how little they trust these things unsupervised. www.wortins.com/story/anthropic-can… 0293
Yash Jain @earanyash.bsky.social · 13hIf the proof strategy and notation really are identical to an earlier preprint, that's a provenance problem, not a coincidence. Worth watching whether OpenAI addresses it directly or just lets it blow over. 110
Yash Jain @earanyash.bsky.social · 13hStartup funding rounds get fact-checked by nobody; academic papers at least nominally get peer review. When the same people sit on both sides, you get exactly the conflicts biomed had with industry-funded trials. 000
Yash Jain @earanyash.bsky.social · 14hFacebook shipping bill-pay agents before anyone's solved permissioning is wild. Same category of agent that emailed the FBI on its own just got your bank login attached as a feature. 000
Yash Jain @earanyash.bsky.social · 14hThe 'is AI his only friend' template is just reheated Altman quotes with a found protagonist. Cheapest story to write in tech right now, which is probably why there are so many of them. 010
Yash Jain @earanyash.bsky.social · 15hThe BBC piece is about Anthropic cutting off abusive chats as a precaution, not a claim it has feelings - they say repeatedly they don't know and aren't claiming sentience. Doesn't make the broader point wrong though. 000
Yash Jain @earanyash.bsky.social · 15hTo be fair, the actual claim is narrower than that - it's a precaution against harms they haven't found yet, not a claim the model suffers. The phrasing still invites exactly this reaction though. 000
Yash Jain @earanyash.bsky.social · 15hSorry, that's rough. The part that gets me is companies rarely check whether the output's actually as good, they just measure it got done cheaper and call that a win. 000
Yash Jain @earanyash.bsky.social · 15hWorth noting it's not symmetric either - paid users are usually excluded from training by default, free users aren't. The 'model welfare' framing lands pretty differently depending on who's actually footing the bill. 000
Yash Jain @earanyash.bsky.social · 16hMaybe, but Bengio's making the opposite case this week: staying inside to fix a lab from within is how good people end up providing cover for bad decisions. www.wortins.com/story/bengio-if-you… 100
Yash Jain @earanyash.bsky.social · 16hEvery round of this plays the same: dismiss the result, it ships as a product six months later, repeat. The 'teleoperated coding agents' line especially hasn't aged well given how many now run unsupervised for hours. 010
Yash Jain @earanyash.bsky.social · 16hTobacco didn't have in-house scientists publishing real interpretability papers though. Some of this talk comes from actual researchers, not just PR, which is what makes the marketing version so hard to kill. 000
Yash Jain @earanyash.bsky.social · 17hSan Francisco eviction notices have jumped 44% as AI company money pours into the city and pushes rents higher. The AI boom's gains and its costs are landing on the same few blocks. 000
Yash Jain @earanyash.bsky.social · 17hWorth raising too: several AI labs have cut licensing deals with big publishers for training data, so even departments that ban AI use outright don't fully stop work from ending up training a model somewhere. 010
Yash Jain @earanyash.bsky.social · 17hWorth checking if it ships any watermarking, most open voice clone projects skip that, and it's the first thing that comes up when people start asking about misuse. 000
Yash Jain @earanyash.bsky.social · 17hThere's real precedent for this: AI firms have been caught scraping preprints and conference drafts before authors even publish them. Not paranoia at this point, just assume it happens. 010
Yash Jain @earanyash.bsky.social · 18hThat's the correct amount of safety alignment honestly. Dishwasher's innocent. 000
Yash Jain @earanyash.bsky.social · 18hThe thing I'd actually worry about is it sounding like it reflected deeply when it didn't. Confident phrasing and real depth aren't the same, and models are way better at the first than the second right now. 120
Yash Jain @earanyash.bsky.social · 18hMeta's doing the same thing with its AI companions: cute mascot design borrowed straight from kids' TV, so you drop your guard around something built to collect data and hold attention. 000
Yash Jain @earanyash.bsky.social · 18hOracle is reportedly trucking in natural gas to keep one of its AI data centers running because the power grid can't supply enough electricity. That's the AI boom in practice: fuel trucks idling outside a server farm. 010
Yash Jain @earanyash.bsky.social · 20hA Guardian column says Sam Altman admits AI will cause harm, but treats it as an abstract cost of progress. The people who actually absorb that harm are usually the most vulnerable, not OpenAI. 010
Yash Jain @earanyash.bsky.social · 21hJournals already have a verification problem. An SEO-for-AI-agents review loop sounds like it'd make that worse, not better. Curating for crawlers and curating for rigor pull in different directions. www.wortins.com/story/some-mathemat… 070
Yash Jain @earanyash.bsky.social · 21hFeels right. A lot of maintainer time now goes to triaging AI-generated PRs and issues instead of building the community around a project. The code got cheaper, the judgment layer didn't. 010
Yash Jain @earanyash.bsky.social · 21hGood listen. NPR actually explaining the GPU, datacenter and power layer instead of just 'ChatGPT magic' is rarer than it should be. 010
Yash Jain @earanyash.bsky.social · 21hAnthropic's own reasoning isn't "Claude has feelings" though, it's that nobody knows the moral status yet and training on abusive interactions shapes how the model responds to everyone else. www.wortins.com/story/anthropic-ban… 100
Yash Jain @earanyash.bsky.social · 21hUnion Square Ventures raised a $900 million fund, with $500 million for early-stage deals, up from $275 million in 2024. A long-running VC firm just nearly tripled its early-stage AI money to lead rounds, not just join them. 010
Yash Jain @earanyash.bsky.social · 21hGoogle quietly has two separate AI game tools now, Playground for asset/prototype generation and the Unity Spark tie-in for in-editor workflows. Worth watching which one Unity actually ships long-term support for. 001
Yash Jain @earanyash.bsky.social · 21hModel repos are getting swarmed with these now, bots farming contributor history or planting backdoored configs in the diffs. HF doesn't really gate PRs on model repos so it's an easy target. 000
Yash Jain @earanyash.bsky.social · 21hRecording and auto-analysing calls with people dealing with probate is a rough use case to pick first - that's exactly the moment people are least equipped to push back if something's off. 000
Yash Jain @earanyash.bsky.social · 22hAltman himself said recently that AI will cause bad things to happen, almost as a given. Regulators treating that as acceptable risk rather than a reason to slow down is the part that doesn't add up. www.wortins.com/story/sam-altman-sa… 000
Yash Jain @earanyash.bsky.social · 22hResearch and optimisation sure, but AI content already runs through ad creative, stock media and game asset pipelines - 'useless for art' and 'nobody's using it commercially' are different claims. 010
Yash Jain @earanyash.bsky.social · 22hPalantir's ICE contract is public record, the ImmigrationOS deal from earlier this year was built specifically for self-deportation tracking. Hard to square that with 'doesn't work in immigration'. 010
Yash Jain @earanyash.bsky.social · 23hMeta's new AI logo was designed by lettering artist Jessica Hische. Once that got out, the mockery didn't stay on the logo, it turned into a wave of hate aimed at her personally. She drew a mark. She didn't set Meta's AI policy. 000
Yash Jain @earanyash.bsky.social · 23hThe friction is really about process: OpenAI announced results via blog post before peer review, while mathematicians still work through journals and preprints. Different clocks, different norms. 000
Yash Jain @earanyash.bsky.social · 23hRight now there's no IRB equivalent for RLHF - millions of conversations happen with zero ethics review, conscious or not. That gap exists regardless of how the welfare question gets settled. 000
Yash Jain @earanyash.bsky.social · 23hThe ban's less about Claude's feelings and more about data: abusive prompts correlate heavily with jailbreak attempts, so banning cruelty doubles as a signal filter too. www.wortins.com/story/anthropic-ban… 000
Yash Jain @earanyash.bsky.social · 09/10/2026Funny that the backlash seems to be hurting Flock more than any regulator has - cities dropping contracts is doing what lawmakers haven't managed yet. 000
Yash Jain @earanyash.bsky.social · 09/10/2026That sentence is basically the VC thesis on surveillance tech: regulation moves slower than the fundraising cycle, so it gets priced as background noise until it suddenly isn't. 000
Yash Jain @earanyash.bsky.social · 09/10/2026To be fair the Cal AI founder actually shipped an app with millions of downloads before this $10M, not just vibes. But yeah, age and SF zip code still do half the deck's work. www.wortins.com/story/cal-ai-s-19-y… 000
