An AI reads raw neural network weights and fixes a small Qwen model. Real breakthrough or hype?
What the article says
- A frontier model called Astra can look at a small network's raw numbers and tell what it was trained to do, such as which game it plays.
- It can also write working game-playing networks from scratch, some beating ones trained the normal way for millions of steps.
- On a small Qwen model, it found the cause of endless reasoning loops and edited a tiny slice of the weights. Maths scores jumped from 8% to 48%.
- The same trick failed on a model that did not loop. Editing can remove an obstacle, but it cannot add skill the model lacks.
- The authors say this could change how models are audited and improved, including self-improvement.
What HN is saying
- A researcher in the field says to calm down. Weight editing has existed for years, and the Qwen fix is basically a well-known steering technique. thread ↗
- Several people point out Astra did not teach Qwen maths. It already knew it, and the edit just stopped it looping. thread ↗
- Others argue the real surprise is a general model doing this at all, and that it may simply have been trained on interpretability research. thread ↗
- A commenter notes the two results get blended. The game-network reading used no tools, while the Qwen fix used a specialised harness. thread ↗
- Replication attempts were mixed. One person got no net gain on a bash-command model, another could not get it to identify their own network. thread ↗
A serious Linux kernel bug is still unpatched on Ubuntu while Debian fixed it weeks ago.
What the article says
- A critical flaw in the Linux kernel's network code has existed for years and can be reached remotely.
- Debian shipped a fix in mid October. Ubuntu, even its long-term releases, was still listed as vulnerable two weeks later.
- The author says Ubuntu promises timely patches for critical bugs and has not delivered, even though it sells security subscriptions.
- He took his servers offline and is moving off Ubuntu, noting that volunteer-run distros like Debian, Arch and Alpine already shipped the fix.
What HN is saying
- Many commenters say this is unacceptable for a company that sells paid security support, and several announced they are switching to Debian or Rocky Linux. thread ↗
- A sharp split on how bad it really is. Some say the official ratings show mostly a crash risk, not a takeover, and that nobody has reported mass attacks. thread ↗
- Others reply that the attack needs no special access, the exploit is out there, and the danger grows each day the patch is missing. thread ↗
- A useful correction: patched kernels are already in Ubuntu's proposed repository, and a firewall rule can block the risky packets in the meantime. Others say a mitigation shouldn't be needed for this severity. thread ↗
- Red Hat was also slow, and Ubuntu's paid Pro customers reportedly did not get the fix either, which undercuts the idea that paying buys speed. thread ↗
Every smart TV spies on you, so what do privacy nerds actually buy instead?
What the article says
- This is a self-post, so these points are inferred from the question and the thread.
- The asker wants a big 4K TV that respects privacy and will pay extra for it.
- They plan to use an Apple TV for streaming and want the TV to act purely as a display.
- They ask whether some TVs pair better with an Apple TV for privacy than others.
What HN is saying
- The most common advice is to buy any TV, never connect it to the internet, and plug in a trusted box over HDMI. thread ↗
- Some say that is not enough, since TVs may join open neighbour WiFi or a guest could connect them. Opening the case to unplug the wireless radio is a popular fix. thread ↗
- Many suggest dumb commercial displays or big monitors, but others warn that even monitors now ship with smart junk and accounts. thread ↗
- Fans of rooting an LG OLED or stripping a TCL with adb say you can keep top picture quality and still own the device. Others call that excessive. thread ↗
- Projectors come up as a naturally dumb option, though daytime brightness is the big complaint. thread ↗
- One poster notes there is no evidence yet of TVs secretly hopping onto neighbours' networks, so staying offline seems to work. thread ↗
Smart glasses that read your silent mouthing with radio waves and turn it into text
What the article says
- Two Georgia Tech students built glasses that let you dictate to your computer without making a sound.
- Radio waves pass through your face and shift as your tongue and lips move. A trained model turns those shifts into text.
- Their first try used electrodes on the face. It failed because of sweat and interference, and nobody wants wires on their face anyway.
- Making it work for everyone meant paying lots of people to read prompts while wearing the glasses.
- Preorders are open for 500 units, with shipping planned for early 2027.
What HN is saying
- Many commenters love the idea and see it as sci-fi made real, from Ender's Game to Her. Some also hear echoes of 1984. thread ↗
- Several people want the tech to clip onto their existing glasses, since prescriptions and big heads make frames a problem. The founder says glasses give them the antenna space and a predictable fit. thread ↗
- Is silent mouthing really easier than whispering? Some commenters found it awkward and self-conscious to try. thread ↗
- Privacy worries came up: one person fears it could be used to spy on people's speech, another says Google dropped similar work over what it might reveal. thread ↗
- Critics say the homepage never explains that this is glasses, and that a sales pitch is a poor fit for a Show HN. thread ↗
Lenovo's fingerprint-locked 'encrypted' USB stick turns out to be a plain SD card in a case.
What the article says
- The page itself was blocked by a bot check, so this is inferred from the title and comments.
- Lenovo sold a fingerprint USB drive as encrypted, but inside it is an ordinary SD card.
- The fingerprint reader only switches the USB connection on and off. The data is never scrambled.
- Anyone can open the case, pull out the card and read it in a normal card reader.
- The flaw was found in 2025 and has an official vulnerability number, but the drive was sold mainly in China.
What HN is saying
- Most people see this as less a bug than a product that never did what it claimed. Several call it plain fraud and ask why nobody faces consequences. thread ↗
- Many say Lenovo had already lost their trust, pointing to the old Superfish adware scandal. thread ↗
- A sharp split on hardware encryption. One side says to just use free software encryption on a normal drive. The other says done properly, hardware can resist brute-force guessing better. thread ↗
- A fingerprint lock is a weak idea anyway, since anyone holding the drive probably has your prints on it. thread ↗
- People struggled to find any product listing, until one commenter found it on Lenovo's China store, which explains why the news is late. thread ↗
- For people wanting a genuinely secure stick, one commenter suggests an open-source drive with hidden encrypted volumes. thread ↗