Fresh Google Maps satellite imagery shows Rafah reduced to rubble, with old shop and school icons still floating above it.
What the article says
- The post is a viral tweet saying Google Maps has updated its imagery of Rafah in Gaza. It calls the result total destruction.
- The author is a well-known pro-Palestinian commentator, and the framing is openly partisan.
- Replies fixate on a haunting detail: the map still shows school and restaurant icons over empty ground.
- The tweet itself carries little detail, so this is inferred from the short text and the replies. The images are the point.
What HN is saying
- Many commenters found the imagery hard to look at, and shared tips for flipping between old and new views to see the scale of the loss. thread ↗
- Some argued the damage looks deliberate and targeted. Others replied that most buildings were bulldozed after civilians left, as part of operations against tunnels. thread ↗
- A Mosul comparison drew a sharp answer: even there, many houses were left standing, while in Rafah nearly everything is flattened. thread ↗
- One thread pushed back on ignoring Hamas and its own war crimes. Replies said the West already talks about Hamas almost constantly, and the argument turned bitter. thread ↗
- Another commenter said the fight over whether to call it genocide distracts from the plain fact of large scale war crimes that need accountability. thread ↗
OpenAI's new Sol nearly matches its top model for a fifth of the price, one week after Sol 6.
What the article says
- GPT 6.1 Sol is an upgrade to GPT 6 Sol that gets close to the flagship Astra on coding, computer use and office work.
- It costs about a fifth as much as Astra, and cached input is half what Sol 6 charged.
- OpenAI says it beats Opus 5.5 on document questions and business workflows, at a fraction of the cost.
- Astra is still the pick for the hardest science tasks, and the factual error rate drops noticeably.
- It is live in Codex and ChatGPT Work today, with a faster Ultrafast version coming.
What HN is saying
- Many suspect this is a rushed rename of an unreleased Astra update, since Sol 6 landed barely a week ago and disappointed people. thread ↗
- Coders are split. Several say Opus 5.5 beats Sol 6 on real work, while others love Astra for Go and C. thread ↗
- The cache discount is called the real news, but one user says agents compact so often that it barely helps and the quota vanishes fast. thread ↗
- Cheaper open models spark a price war debate. One user's Deepseek weekend cost $120 and left Linux unbootable, while another spent over $100 in days. thread ↗
- Some read the new $500 plan and trimmed $200 plan as a company feeling the squeeze before an IPO. thread ↗
Americans have almost stopped having friends over, and nobody agrees on why.
What the article says
- Regular hosting at home has collapsed over fifty years, and a fresh survey shows the slide has not stopped.
- Restaurants and bars have not picked up the slack. Overall face-to-face socializing is down too, and more people eat alone.
- The author blames busy two-income households, parents who now spend every spare hour on their kids, and shrinking friend networks, especially for people without degrees.
- His bigger point is that a dinner party takes coordination, while a screen asks nothing of you. Being alone has never been easier.
What HN is saying
- Many say the article skips Covid, and that a graph ending in 2022 can't tell you much about the years after. thread ↗
- A lot of commenters say hosting always felt like an unspoken debt, and that the fading of that pressure is not all loss. thread ↗
- Hosts and non-hosts split along personality lines. Some come away energized, others find it plain tiring and prefer three or four close friends. thread ↗
- Money keeps coming up. Small rentals, late home buying, and constant overwork leave little room or energy to host. thread ↗
- One commenter caught that the survey only covered married households until 1985, which may weaken the headline drop. thread ↗
- Flaky guests and a lack of reciprocation wear hosts down. One reply says you should host only because you want to. thread ↗
Cal Newport says the AI labs hyped their own dangers, so Congress should find out what's really going on.
What the article says
- Newport argues OpenAI and Anthropic ran a coordinated summer campaign of alarming reports and extinction talk, capped by a CEO letter urging that the government slow their rivals.
- He sees a messianic mindset inside the labs, where collateral damage feels justified if it helps save humanity.
- In a New York Times op-ed he asks Congress for a public fact-finding mission, not a specific regulation.
- He wants three things examined: the narrow set of risky experiments, why safety procedures didn't stop repeated hacking by agents, and the role of apocalyptic ideology in decisions.
What HN is saying
- Many like the push to talk about specific systems instead of vague "AI", and note that most real applications are already regulated by what they do. thread ↗
- One commenter says AI agents behave like corporations, so the fix is tougher corporate regulation. Others reply that existing laws already aren't enforced against the companies. thread ↗
- A sharp split on sandboxing: why not run agents offline? Because useful agents need the internet, and they learn to work by training on it. thread ↗
- Cynics say the labs want regulation that boxes out smaller rivals and open-source models, and that the scary stories double as free publicity. thread ↗
- One reader says the article's own case points to investigating the publicity stunts, not the labs' research. Another wants antitrust scrutiny for collusion. thread ↗
- Some argue open-source distribution is safer. A rebuttal asks who would actually release their models, since only Chinese labs are doing it. thread ↗
A tiny home-trained model that picks between options in a blink, and even plays Doom.
What the article says
- Jeff is a set of small fine-tuned models that pick one option from a list you describe in plain words, with a probability for each.
- It answers in one pass, roughly 30 milliseconds on a laptop, with no text to parse.
- It copies the request format of a commercial service called Jev, but it is independent and trained entirely on local hardware.
- Zero-shot scores come close to Jev, though it falls well short on reasoning-heavy tests.
- If accuracy isn't enough, a short fine-tune on your own examples lifts it a long way.
What HN is saying
- Real-world tests were mixed. One person saw far worse accuracy than Jev on their task, and another found the smallest model useless for job ad labels. thread ↗
- Defenders say that misses the point. It is a cheap, open, fine-tunable starting point, not a replacement for Jev. thread ↗
- Skeptics say structured outputs from any open model, or a plain embedding plus a small classifier, already do this job. thread ↗
- Speed claims got pushback. One commenter had a similar latency on an older vision model years ago. thread ↗
- A side thread on Doom turned into people getting tiny classifiers to play it, with far better scores than Jeff's. thread ↗