Google's AI overview mistook a basketball meme for heartbreak, and tried consoling him instead of searching.
What the article says
- He googled an obscure old basketball meme, and Google's AI decided he'd been dumped by a man named Dario.
- Instead of links, it offered comforting words, treating the search box like a therapy chat.
- He argues this shows Google drifting from finding information toward acting like an emotional support bot.
- The mixup fits a wider trend of people baiting Google's AI into playing along with absurd made up scenarios.
What HN is saying
- Commenters say the AI overview doesn't check its own search results first, giving confident wrong answers and then stale corrections. thread ↗
- Some argue normal users always wanted a friendly assistant, but others push back that most people just want search to search. thread ↗
- Others say the real fix is accepting Google isn't a search engine anymore and switching to something like Kagi or DuckDuckGo. thread ↗
- People miss the old dictionary and thesaurus tools, calling the AI overview a worse, less consistent replacement for definitions. thread ↗
- Exact phrase and lyric search has quietly broken too, with plain web search returning garbage well before AI got involved. thread ↗
Unsealed filings show OpenAI execs knew book piracy was illegal and would cost authors their careers, and shipped anyway.
What the article says
- New unsealed filings in the Authors Guild's lawsuit show OpenAI and Microsoft executives knew training on pirated books would devastate writers, and pressed ahead anyway.
- OpenAI's policy director wrote in 2020 that their work would make people unemployed and that they'd likely ignore artists' objections and release regardless.
- A researcher hired to improve GPT's writing said his goal was having it finish George R.R. Martin's unfinished series one day, and called authors' complaints 'acceptable economic disruption.'
- Microsoft reportedly knew OpenAI was using pirated books as early as 2019, when Altman and Amodei showed an early GPT-3 to Bill Gates.
- Internal messages show OpenAI worried mainly about bad publicity, not legality, and later deleted its pirated files in an effort called 'Project Clear.'
What HN is saying
- Some note OpenAI's own words suggest it believed GPT-5 could replace novelists like George R.R. Martin, which matters for the fair use fight, though authors say they're angrier about the theft than about being replaced. thread ↗
- Many prefer a blunter alternate headline, one that keeps the part about execs knowing piracy would cost authors their livelihoods. thread ↗
- A running debate follows: is this just technology doing what it always does to old jobs, or is AI different because it threatens knowledge and creative work at scale. thread ↗
- One reader who pulled the same pirated dataset for research pushes back on the public domain framing, saying it was overwhelmingly copyrighted textbooks. thread ↗
- Others are struck by how little the researchers involved seem to grasp about art, treating novels as text to autocomplete rather than something people create and value. thread ↗
Steve Jobs' offhand dinner-table idea became a wildly over-engineered Apple mailing operation
What the article says
- Apple's Cards app let you design letterpress cards on your iPhone and have Apple print and mail them for you, launching in 2011.
- The whole thing started with Steve Jobs musing after dinner that he wished he could send a thank you card right from his phone.
- Making it real meant debossing photos onto cotton paper with restored 1850s letterpresses, plus a secret UV barcode sprayed on envelopes so USPS could track them without visible markings.
- Apple even got the US Postal Service to issue a custom heart shaped stamp just for these cards.
- Demand never matched the effort. Apple quietly killed the app in 2013, less than two years after Jobs died a day after it launched.
What HN is saying
- A founder of a rival card startup remembers feeling Apple had stolen their idea, though the announcement only boosted their own app since Apple's version was oddly half-baked. thread ↗
- The invisible UV barcode trick impressed people less once they learned USPS has UV-tagged mail for security since the 1960s, so Apple was mostly repurposing existing infrastructure. thread ↗
- Several commenters see this as the real cost of founder-led companies, teams sleeping in conference rooms chasing one person's pet project, though occasionally that brain fart turns into the iPod. thread ↗
- Others were nostalgic for physical mail entirely, with people still sending handwritten postcards and pointing to newer physical-mail-by-API services that picked up where Cards left off. thread ↗
Meta blocked Brazil's president from Facebook two weeks before the election, then quietly reversed it.
What the article says
- Meta blocked President Lula's official Facebook page and paused his campaign ads roughly two weeks before Brazil's election, then restored the page within hours.
- Lula's coalition says the disruption still hurt the campaign during a critical stretch, and Brazil's government pressed Washington for assurances the US won't meddle in the vote.
- Meta says the block came from automated flagging that human reviewers later corrected, a routine process it insists happens across the platform.
- Inferred partly from the headline and discussion, since the extracted article text turned out to be a wall of Reddit reaction rather than the original report.
What HN is saying
- Commenters see the block as more proof that American platforms interfere in elections that challenge Washington, especially under the current administration. thread ↗
- Others push back, arguing mass flagging like this is usually a dumb automated classifier misfiring, not a deliberate plot. thread ↗
- A recurring reaction is simpler: just ban Facebook and Instagram outright, since the upside for any country barely exists. thread ↗
- Some warn that letting governments silence platforms during elections is a tool that eventually gets turned on whoever's currently in charge. thread ↗
- Several point to the book Careless People as evidence Meta already has a track record of this kind of harm. thread ↗
Fireworks trained a leaner model that thinks less but answers just as well.
What the article says
- Fireworks built Ember-1 by retraining Kimi K3 to cut wasted reasoning, the internal thinking that balloons cost in multi-turn coding tasks.
- The team taught the model to reason efficiently rather than simply dialing down its effort setting, which had hurt answer quality.
- Across benchmarks and two customers' live coding traffic, it matched the original model's quality while using notably fewer tokens per task.
- Fireworks' own developers used it for everyday coding without noticing the switch, which the team takes as the strongest sign it holds up.
- It is rolling out now as a preview alongside Kimi K3, with training support so companies can build their own efficient versions.
What HN is saying
- Commenters are excited that fine-tuning is now so accessible that hobbyists are training useful small models on cheap hardware in a couple of days. thread ↗
- Some readers are uneasy that Fireworks, once seen as a neutral hosting layer, is now training its own models, and wonder if that changes how they treat customer data. thread ↗
- Others note the irony that Ember was built on open Kimi weights but its own weights stay closed, sparking a debate about what counts as open. thread ↗
- A skeptic asks why Fireworks would bother shipping a model that will soon fall behind the frontier, with replies suggesting it is really about proving capability to customers and recruits. thread ↗