03-30 00:01 · AI,技术,HackerNews,AI
Police used AI facial recognition to wrongly arrest TN woman for crimes in ND
A Tennessee grandmother spent more than five months in jail after police used an AI facial recognition tool to link her to crimes committed in North Dakota – a state she says she’d never been to before.
Police in Fargo, North Dakota, have acknowledged “a few errors” in the case and pledged changes in their operations but stopped short of issuing a direct apology.
Angela Lipps, 50, was first arrested in Tennessee on July 14, according to a statement from the Fargo Police Department and a verified GoFundMe for Lipps.
Unbeknownst to Lipps, a warrant had been issued for her arrest weeks earlier – in Fargo, over 1,000 miles away from her Tennessee home. Months before, several instances of bank fraud had occurred in and around Fargo, according to police.
In their search for a suspect in the bank fraud cases, investigators used “our partner agency’s facial recognition technology” as well as “additional investigative steps independent of AI to assist in identification” before submitting the report to the Cass County State Attorney’s Office, Fargo Police Department Chief Dave Zibolski told CNN in an email.
But Zibolski said at a Tuesday news conference that his police department’s reliance on some of the information from a neighboring agency’s AI system is “part of the issue,” referring to errors made in Lipps’ case.
“At some point, our partner agency over at West Fargo purchased their own AI facial recognition system that we were not aware of at the executive level …, and we would not have allowed that to be used, and it has since been prohibited,” he said.
The West Fargo Police Department told CNN that they use Clearview AI, a startup with a database of billions of photos scraped from the internet, including social media. Clearview “identified a potential suspect with similar features to Angela Lipps” and West Fargo police shared that report with Fargo police, reads a statement from the police department. The statement notes that West Fargo police didn’t forward any charges and didn’t have enough evidence to charge anyone for the fraud case in West Fargo.
CNN has reached out to Clearview AI for comment. It’s unclear what other evidence was used in the investigation to tie Lipps to the crimes.
Lipps’ case comes as police departments across the country have rapidly integrated new technologies, including AI. But police use of the novel technology has attracted criticism – and it’s been linked to other cases of misidentification.
‘Terrified and exhausted and humiliated’
On July 1, a North Dakota judge signed a warrant for Lipps’ arrest, with nationwide extradition. She was arrested July 14 and spent over three months in a Tennessee jail before being extradited, according to Fargo police and her lawyers.
It wasn’t until October that Tennessee law enforcement told the Cass County Sheriff’s Office in North Dakota they had Lipps’ extradition waiver. She was facing multiple charges, including felony theft and felony unauthorized use of personal identifying information, according to her lawyers.
It’s unclear why it took so long for Tennessee authorities to notify their North Dakota counterparts about Lipps’ arrest. Lipps’ attorneys told CNN they “have seen a July 14, 2025 email notifying various North Dakota law enforcement personnel that Angela had been arrested in Tennessee.”
Fargo police, alternately, told CNN, “We have been unable to determine based on available information if the length of time Ms. Lipps was in jail in Tennessee before being transported to North Dakota was due to serving time for a probation violation or if it was because she fought extradition.”
CNN has reached out to Tennessee authorities for comment.
Lipps’ extradition to North Dakota, she said in her GoFundMe, was terrifying: “It was the first time I had ever been on an airplane,” she wrote. “I was terrified and exhausted and humiliated.”
In Fargo, she was given a lawyer who found bank records showing she had been in Tennessee during the time of the crimes, according to the GoFundMe. Fargo police say on December 12, the State’s Attorney’s Office informed the Fargo detective that the defense had produced “potential exculpatory evidence.”
On December 23, the Fargo detective, the state’s attorney and the judge “mutually agreed to dismiss the charges without prejudice to allow for further investigation,” according to Fargo police. Lipps was released from custody on Christmas Eve.
For Lipps, her months of incarceration were devastating.
“The trauma, loss of liberty, and reputational damage cannot be easily fixed,” Lipps’ lawyers told CNN in an email. Her lawyers said Lipps was unavailable to speak for an interview.
Lipps, a mother of three and grandmother of five, had never been to North Dakota before her extradition, according to CNN affiliate WDAY.
And after her ordeal, she never plans to return to the state: “I’m just glad it’s over,” she told WDAY. “I’ll never go back to North Dakota.”
Her legal team says they’re investigating why Lipps was held in
▸ 展开全文
03-30 00:01 · AI,技术,HackerNews,AI
Kyushu Railway Company Train Varieties
Train Varieties
Distinctive and sophisticated, JR Kyushu trains are full of character and unique features.
Here, we will introduce three types of trains with different features.
Design & Story Trains
(Sightseeing Trains)JR Kyushu trains are full of character. Noted for their distinctive and sophisticated exteriors and interiors, they are packed with unique features which allow passengers to enjoy Kyushu's breathtaking landscape and passing scenery. Riding Design & Story Trains (sightseeing trains) brings out the best of Kyushu and is an unforgettable experience in itself.
YUFUIN NO MORI
Resort train that leads you to Yufuin, an exquisite hot springs resort.
ASO BOY!
Let' s go see the Aso caldera! The train will take you there as you play with "Kuro".
A-TRAIN
Named after a famous jazz tune, this is a train for adult travelers that transports those on board back to the good old days.
IBUSUKI NO
TAMATEBAKOTry this moving monotone magical treasure box. White smoke billows when you board!
UMISACHI
YAMASACHINichinan Line’s resort express allows you to enjoy the rich natural surroundings of Miyazaki, a land of myth.
KAWASEMI
YAMASEMIExperience the Hitoyoshi/Kuma region like the wild birds that frequent its majestic mountains and crystal clear streams.
TWO STARS 4047
Begins service in fall 2022!
A train to tour the seas of West Kyushu
36+3
(Sanju-Roku plus San)A new journey around Kyushu
ARU RESSHA
A luxury train that has been revived after 100 years
KANPACHI/ICHIROKU
A new sightseeing train to taste the culture of the Yufu Kogen Line
Bento box lunch reservation service
On the D & S trains of JR Kyushu, a variety of bento box lunches and snacks are sold on the train. Please note that some of the items need to be reserved in advance, before boarding the train.
Shinkansen
(Bullet Train)The Kyushu Shinkansen not only reduces travel time around the Kyushu area to stimulate interaction between people as well as economic and cultural exchange - it plays a key role as the pipeline for business and tourism that connects Kyushu with the Kansai and Sanyo areas.
800 Series
ShinkansenA luxurious train trip in a car decorated with traditional Japanese arts and crafts
N700 Series
ShinkansenAerodynamically the fastest train connecting Kyushu and Honshu.
Nishi Kyushu
ShinkansenBegins service in fall 2022!
Please look forward to seeing yet another leap forward for the “Kamome” as the Nishi-Kyushu Shinkansen.
Other Limited Express
The railways operated by JR Kyushu are not just a means of transportation - they connect the cities of Kyushu. Full of personality, simply riding one makes for a fascinating experience in itself. Whether it be a relaxing trip or one that is just for fun, these trains will take you on tranquil journeys through the breathtaking landscapes and nature that make up Kyushu.
Relay KAMOME
Begins service in fall 2022!
SONIC
(883 Series)East Kyushu intercity urban express that is true to its "SONIC" name.
HUIS TEN BOSCH
Direct route to "Huis Ten Bosch", the largest theme park in Japan!
KYUSHU ODAN
TOKKYUTravel to famous sightseeing spots on this red-hot train!
Other limited
express train servicesEach area and each line offers its own unique railroad experience. Try to experience each of the trains.
Train Lineup
Design & Story Trains (Sightseeing Trains)
Shinkansen (Bullet Train)
Other Limited Express
Timetables
This PDF includes the timetables for regular services of trains operated by Kyushu Railway Company.
*JR Kyushu Rail Pass can’t be used on the Shinkansen between Hakata Sta. and Shin-Osaka Sta.
Purchase a Rail Pass
You can purchase JR Kyushu Rail Passes and reserve seats on popular Kyushu trains.
One-way and
round-trip ticketsPurchase one-way and round-trip tickets, reserved seats, and early discount tickets from the JR-KYUSHU Train Reservation page.
▸ 展开全文
03-30 00:01 · AI,技术,HackerNews,AI
Midnight train from GA: A view of America from the tracks as airports struggle
Midnight train from Georgia: A view of America from the tracks as airports struggle in the shutdown
ABOARD THE CRESCENT (AP) — There’s something melodic about watching the sun rise over a rural stillness broken only by the rhythms of steel wheels on tracks. Or so we tell ourselves.
In this case, being aboard a train at all owed more to politics than poetry.
Congress and Donald Trump were mired in their latest budget stalemate, one rooted in the Republican president’s immigration crackdown and the tactics of federal forces he has sent to U.S. cities. But this impasse has upended a foundational constant of American life today: easy air travel.
In Atlanta, my hometown airport, cheerfully marketed as the world’s busiest, had descended into organized chaos. Unpaid federal employees called out from work, leaving a diminished security staff to screen travelers frustrated by hourslong waits in line. I wanted to get to Washington for the NCAA basketball tournament. So I eliminated the risk of a missed flight and booked the train overnight and into game day across a 650-mile route.
In this fraught moment in U.S. politics, I slowed down and thought about things we take for granted. Who ever ponders the conveniences of that 20th-century innovation, the airplane, that makes 21st-century hustle possible? We book and board. An unconscious, first-world flex of modernity. It’s even rarer to grapple with the inconvenience.
My decision had taken me further back, to the 19th century and another defining innovation: the long-distance train.
A 14½-hour weekend train ride is time aplenty to appreciate how completely politics, economics, social strife and fights over identity and belonging have always affected the order of our lives, including how, when and where we move around in these United States. But Amtrak's Crescent also allowed me to see the expanse of our collective experience.
I traversed the urban, suburban and rural breadth of East Coast America. I learned how other travelers came aboard. And in that, I found the portrait of people, past and present, who refuse to be as paralyzed as some of their elected leaders.
There is little glamour late night in a crowded Amtrak station. Children are up past bedtime and tended by frazzled parents. Older adults struggle with luggage and stairs.
Airports are not red-carpet affairs either, of course. But there is a certain cache to Delta's Atlanta-Washington flights. They typically take about two hours gate to gate. They often are slotted at a midpoint gate of the concourse nearest the main terminal. That is almost certainly a nod to members of Congress who use it — but who have lost some airline perks during this extended partial shutdown.
In normal circumstances I can get from my front porch to Capitol Hill or downtown in as little as 4½ hours. Security lines these days could at least double my overall air travel time.
The train is still longer, and time is money, we are taught. But certainty has value, too, even if it means at 11:29 p.m. departure. And at the Amtrak station, there were no standstill lines, no Transportation Security Administration agents, no ICE agents as stand-ins.
Passengers who arrived mere minutes before departure made it on board and found seats quickly — assigned in boarding order, not predetermined zones that yield jammed aisles. There’s no in-seat service or satellite TV. But even coach seats, the lowest Amtrak tier, are as spacious as airline first-class – and there is Wi-Fi, so it's not the 19th century or even 20th century after all.
On board, I heard one crew member joke, “I'm no TSA agent.”
As a boy in rural Alabama, I counted train cars and wondered where they were headed. I’ve since read diary entries and letters from my grandmother and her sisters recounting World War II-era weekend trips to Atlanta.
The South's largest city has a historical hook, too. Originally named “Terminus,” Atlanta developed in the antebellum era as a critical intersection of north-south and east-west rail routes. That is what drew Gen. William Tecumseh Sherman for one of the Civil War’s seminal campaigns that helped defeat the Confederacy.
A century after the Civil War, Delta chose Atlanta for its headquarters rather than Birmingham, Alabama, which was the larger city as of the 1960 census. The company's decision was tied up in tax breaks for the airline, named for its crop duster origins in the Mississippi Delta region. According to some interpretations, Delta's decision was made easier because of the more overt racism of Alabama's and Birmingham's leaders as they defended Jim Crow — a code that, among other acts, allowed states to segregate the passenger trains that predated Amtrak.
On this night, I heard many languages and accents, notable given the role that immigrant labor played in building the U.S. rail system and especially striking now with immigration — legal and illegal — at the forefront in Washington, my destination. I saw faces that reflected U.S. pluralism, a
▸ 展开全文
03-30 00:01 · AI,技术,HackerNews,AI
Coding Agents Could Make Free Software Matter Again
I’ve been vibe-coding a lot lately. Like, a lot a lot. Maybe not quite the “AI psychosis” Andrej Karpathy recently joked about on No Priors, but not wildly far off either.[1] But the more I vibe, the more a thought recurs, i.e. that AI coding agents may be about to make free software matter more than it ever has. Not open source in the bland corporate sense. I mean free software in Stallman’s sense: software that gives users the freedom to run it, study it, modify it, and share it.
Even for the relatively few who were aware of the distinction, it has felt mostly academic for a long time. SaaS made it hard to care about software freedom because most people never saw or touched the source code of software they depended on in the first place. The code lived on someone else’s servers, the vendor handled operations, and the practical question became convenience, not freedom.
Agents change that. If an agent can read a codebase, understand it, and modify it on your behalf, then access to source code stops being a symbolic right for programmers and becomes a practical capability for far more people. Suddenly the difference between software you can change and software you can only beg starts to really matter.
And I don’t just think this in the abstract. I recently tried to get an AI agent to customize a SaaS app for me, and the experience made the whole problem very concrete very fast.
Free software once mattered deeply, but faded when SaaS made those freedoms feel irrelevant.
In 1980, Richard Stallman[2] was a programmer at MIT’s AI Lab, and he had a problem with a printer. The lab had gotten a new Xerox laser printer, and it kept jamming. Stallman wanted to fix the issue — or at least add a feature to notify users when their print jobs got stuck — but Xerox wouldn’t give him the source code. The printer’s software was proprietary.
This seems like a small thing. It was not a small thing to Stallman.
He’d grown up in a computing culture where sharing code was the norm. When you got software, you got the source code, because of course you did — how else would you fix bugs or add features? The Xerox printer incident crystallized something for him: a world where software was locked up, where you couldn’t study or modify the tools you depended on, was a world where users had lost something fundamental.
So Stallman founded the Free Software Foundation and spent the next four decades evangelizing what he called “the four freedoms”:
- Freedom 0: The freedom to run the program as you wish, for any purpose.
- Freedom 1: The freedom to study how the program works, and change it to do what you want.
- Freedom 2: The freedom to redistribute copies so you can help others.
- Freedom 3: The freedom to distribute copies of your modified versions to others.
“Free as in speech,” he’d say, “not free as in beer.”
For a while, this message resonated. The 1990s saw an explosion of free software: Linux, Apache, MySQL, PHP — the entire stack that would come to power most of the internet. Companies like Red Hat proved you could build real businesses around it. Eric Raymond wrote “The Cathedral and the Bazaar” and argued that open development produced better software. Microsoft’s Steve Ballmer called Linux “a cancer.” It felt like a genuine ideological battle for the soul of computing.
And then, quietly, the battle became irrelevant, for a pretty boring reason.
But we’ll come back to that in a moment.
The “open source” rebrand preserved code sharing while stripping out the user-rights philosophy.
First, here’s a piece of history that I didn’t know before researching this post. It matters for understanding what’s happening now.
On February 3, 1998, a group of people met at the Foresight Institute in Palo Alto — not a software organization, but a nanotechnology think tank. Christine Peterson, the institute’s executive director, proposed replacing “free software” with a new term: “open source.”[3] Her reasoning was practical: every time you said “free software,” people thought you meant free-as-in-beer, and you’d spend ten minutes explaining the difference instead of talking about the actual software.
A few weeks later, at Tim O’Reilly’s April 1998 “Freeware Summit,” attendees debated naming and voted 9-6 for “open source” over alternatives like “sourceware” and “free software.”[4]
The important thing is what got lost in the rebrand. Eric Raymond and Bruce Perens co-founded the Open Source Initiative that same month, and Raymond published a manifesto called “Goodbye, ‘free software’; hello, ‘open source.'” His key argument: the old terminology made corporate types nervous.[5]
Stallman was not invited to Tim O’Reilly’s “Freeware Summit” in April 1998 — the event that helped cement the new leadership narrative. Linus Torvalds was invited. Larry Wall was invited. Guido van Rossum was invited. Stallman was not.[4]
Why does this matter? Because the “open source” rebrand wasn’t just a marketing change — it was a philosophical amputation. “Open source
▸ 展开全文
03-30 00:01 · AI,技术,HackerNews,大模型,AI
ChatGPT won't let you type until Cloudflare reads your React state
Edit April 2, 2026: I've been getting inbound interest from researchers wanting to run their own queries. The MCP integration I use for my own research lets you analyze live mobile telemetry continuously collected from real devices in the wild, directly from Claude. To access it reach out at buchodi@proton.me
Every ChatGPT message triggers a Cloudflare Turnstile program that runs silently in your browser. I decrypted 377 of these programs from network traffic and found something that goes beyond standard browser fingerprinting.
The program checks 55 properties spanning three layers: your browser (GPU, screen, fonts), the Cloudflare network (your city, your IP, your region from edge headers), and the ChatGPT React application itself (__reactRouterContext
, loaderData
, clientBootstrap
). Turnstile doesn't just verify that you're running a real browser. It verifies that you're running a real browser that has fully booted a specific React application.
A bot that spoofs browser fingerprints but doesn't render the actual ChatGPT SPA will fail.
The Encryption Was Supposed to Hide This
The Turnstile bytecode arrives encrypted. The server sends a field called turnstile.dx
in the prepare response: 28,000 characters of base64 that change on every request.
The outer layer is XOR'd with the p
token from the prepare request. Both travel in the same HTTP exchange, so decrypting it is straightforward:
outer = json.loads(bytes(
base64decode(dx)[i] ^ p_token[i % len(p_token)]
for i in range(len(base64decode(dx)))
))
# → 89 VM instructions
Inside those 89 instructions, there is a 19KB encrypted blob containing the actual fingerprinting program. This inner blob uses a different XOR key that is not the p
token.
Initially I assumed this key was derived from performance.now()
and was truly ephemeral. Then I looked at the bytecode more carefully and found the key sitting in the instructions:
[41.02, 0.3, 22.58, 12.96, 97.35]
The last argument, 97.35
, is the XOR key. A float literal, generated by the server, embedded in the bytecode it sent to the browser. I verified this across 50 requests. Every time, the float from the instruction decrypts the inner blob to valid JSON. 50 out of 50.
The full decryption chain requires nothing beyond the HTTP request and response:
1. Read p from prepare request
2. Read turnstile.dx from prepare response
3. XOR(base64decode(dx), p) → outer bytecode
4. Find the 5-arg instruction after the 19KB blob → last arg is the key
5. XOR(base64decode(blob), str(key)) → inner program (417-580 VM instructions)
The key is in the payload.
What the Decrypted Program Checks
Each inner program uses a custom VM with 28 opcodes (ADD, XOR, CALL, BTOA, RESOLVE, BIND_METHOD, JSON_STRINGIFY, etc.) and randomized float register addresses that change per request. I mapped the opcodes from the SDK source (sdk.js
, 1,411 lines, deobfuscated).
The program collects 55 properties. No variation across 377 samples. All 55, every time, organized into three layers:
Layer 1: Browser Fingerprint
WebGL (8 properties): UNMASKED_VENDOR_WEBGL
, UNMASKED_RENDERER_WEBGL
, WEBGL_debug_renderer_info
, getExtension
, getParameter
, getContext
, canvas
, webgl
Screen (8): colorDepth
, pixelDepth
, width
, height
, availWidth
, availHeight
, availLeft
, availTop
Hardware (5): hardwareConcurrency
, deviceMemory
, maxTouchPoints
, platform
, vendor
Font measurement (4): fontFamily
, fontSize
, getBoundingClientRect
, innerText
. Creates a hidden div, sets a font, measures rendered text dimensions, removes the element.
DOM probing (8): createElement
, appendChild
, removeChild
, div
, style
, position
, visibility
, ariaHidden
Storage (5): storage
, quota
, estimate
, setItem
, usage
. Also writes the fingerprint to localStorage
under key 6f376b6560133c2c
for persistence across page loads.
Layer 2: Cloudflare Network
Edge headers (5): cfIpCity
, cfIpLatitude
, cfIpLongitude
, cfConnectingIp
, userRegion
These are injected server-side by Cloudflare's edge. They exist only if the request passed through Cloudflare's network. A bot making direct requests to the origin server or running behind a non-Cloudflare proxy will produce missing or inconsistent values.
Layer 3: Application State
React internals (3): __reactRouterContext
, loaderData
, clientBootstrap
This is the part that matters. __reactRouterContext
is an internal data structure that React Router v6+ attaches to the DOM. loaderData
contains the route loader results. clientBootstrap
is specific to ChatGPT's SSR hydration.
These properties only exist if the ChatGPT React application has fully rendered and hydrated. A headless browser that loads the HTML but doesn't execute the JavaScript bundle won't have them. A bot framework that stubs out browser APIs but doesn't actually run React won't have them.
This is bot detection at the application layer, not the browser layer.
The Exit: How the Token Is Built
After collecting all 55 properties, the program hits a 116-byte encrypted blob that decrypts
▸ 展开全文
03-30 00:01 · AI,技术,HackerNews,大模型,AI
Claude Code runs Git reset –hard origin/main against project repo every 10 mins
My automated tool is running git reset --hard origin/main in my project every 10 minutes #40710
Description
Update (2026-03-30): Root cause identified — this was NOT a Claude Code bug.
The resets were caused by a separate tool I built that was running locally, which used GitPython to hard-reset the working directory on a poll cycle. The evidence was misleading because the tool shared the same CWD and monitored projects CC was working on. See my comment below for full details. Apologies for the false report.
Summary
Claude Code performs git fetch origin
+ git reset --hard origin/main
on the user's project repo every 10 minutes via programmatic git operations (no external git
binary spawned). This silently destroys all uncommitted changes to tracked files. Untracked files survive. Git worktrees are immune.
Environment
- Claude Code version: 2.1.87 (Homebrew cask, compiled Bun binary)
- OS: macOS 15.4 (Darwin 25.3.0, arm64)
- Shell: zsh
Evidence
1. Git reflog: 95+ entries at exact 10-minute intervals
e8ea2c9 HEAD@{2026-03-29 22:19:09 +0200}: reset: moving to origin/main
e8ea2c9 HEAD@{2026-03-29 22:09:09 +0200}: reset: moving to origin/main
e8ea2c9 HEAD@{2026-03-29 21:59:09 +0200}: reset: moving to origin/main
e8ea2c9 HEAD@{2026-03-29 21:49:09 +0200}: reset: moving to origin/main
...
8792b6c HEAD@{2026-03-29 16:55:41 +0200}: reset: moving to origin/main
8792b6c HEAD@{2026-03-29 16:45:41 +0200}: reset: moving to origin/main
...
32aa7c7 HEAD@{2026-03-28 15:47:36 +0100}: reset: moving to origin/main
32aa7c7 HEAD@{2026-03-28 15:37:36 +0100}: reset: moving to origin/main
The second offset is consistent within each session but varies between sessions (:08
, :36
, :41
, :09
), confirming a timer tied to session start time with a 600-second interval. 95+ entries observed across 4 sessions over ~36 hours.
2. Live reproduction
- Modified
src/lib/api.ts
(tracked file) and created.canary-test.txt
(untracked file) - Monitored every 15 seconds
- At the next 10-minute mark,
api.ts
silently reverted — modification gone .canary-test.txt
(untracked) survived- Reproduced consistently across 4 consecutive cycles
3. fswatch caught the file operations
At the exact reset time, fswatch on .git/
captured:
23:59:10.349 .git/refs/remotes/origin/HEAD.lock Created IsFile Removed AttributeModified
23:59:10.352 .git/logs/HEAD IsFile Updated
23:59:10.354 .git/refs/heads/main.lock Created IsFile Removed AttributeModified
This is the classic pattern for git fetch origin
+ git reset --hard origin/main
.
4. Only the Claude Code process is a candidate
lsof
confirms the Claude Code CLI process (PID 70111, claude --dangerously-skip-permissions
) is the only process with CWD in the affected repo. Two other Claude CLI sessions are in different directories.
5. No external git binary spawned
Process monitoring at 0.1-second intervals found zero git
processes around reset times. The operations are programmatic (libgit2 or similar) within the Claude Code process, confirmed by .git/
lock file creation without any external process.
6. Worktrees are immune
The worktree reflog shows zero reset: moving to origin
entries. The reset targets the main working tree only.
What was ruled out
A thorough investigation eliminated all external causes:
Binary analysis (partial)
From the compiled binary at /opt/homebrew/Caskroom/claude-code/2.1.87/claude
:
hg1()
function does["fetch","origin"]
viat_(C8(), _)
without explicit CWD, defaulting toprocess.cwd()
io1()
function is a git pull wrapper logginggit pull: cwd=${H} ref=${_??"default"}
fileHistory
state tracks{snapshots: [], trackedFiles: new Set, snapshotSequence: 0}
- The exact timer/setInterval setup could not be identified in minified code
Impact
Any uncommitted changes to tracked files in the main working tree are silently destroyed every 10 minutes. During a 2-hour session, changes had to be re-applied 3+ times before the cause was identified. The bug is invisible when all changes are committed (the reset is a no-op), making it appear intermittent.
Question for the Claude Code team
What internal mechanism runs git fetch origin
+ git reset --hard
against process.cwd()
every 600 seconds? This could not be determined from the outside due to the compiled binary and lack of sudo for process tracing.
Workarounds
- Use git worktrees — confirmed immune (zero reset entries in worktree reflog)
- Commit frequently — committed changes survive the reset
Related issues
- Critical Bug: Code Revisions Being Repeatedly Reverted #8072 — "Critical Bug: Code Revisions Being Repeatedly Reverted"
- [MODEL] CRITICAL: Claude executed git reset --hard without authorization causing data destruction #7232 — "CRITICAL: Claude executed git reset --hard without authorization causing data destruction"
- [BUG]
claude install
corrupts project git remote URL to anthropics/claude-plugins-official on macOS #32793 — "claude install
corrupts project git remote URL" (marketplace git in wrong directory — related but not this specific bug)
▸ 展开全文
03-29 11:25 · AI,技术,HackerNews,AI
What if AI doesn't need more RAM but better math?
@adlrocha - What if AI doesn’t need more RAM but better math?
How TurboQuant compresses the KV cache without losing accuracy, and what that could mean for memory stocks
Last week I was writing about the hardware side of the AI memory problem: the HBM density penalty, the EUV bottleneck, and the supply chain pressure squeezing DRAM prices for everyone from data centre operators down to consumer electronics. This week, Google published something that attacks the exact same problem using another approach: not “build more memory”, but “need less of it.”
You guessed it! This post will dive a bit deeper into what TurboQuant is, and what this may imply to the field of AI. What Pied Piper achieved in the Silicon Valley TV Show with their general-purpose lossless compression algorithm, Google may have achieved it for the compression of information represented as vectors in a high-dimensional space.
What is a transformer? And the KV cache?
But before getting into what TurboQuant does, let’s make a brief detour to understand what is this algorithm is actually built to compress, and why it is important for LLMs and the memory problem.
GPT models are what are known as autoregressive: they generate text one token at a time, where each new token is conditioned on everything that came before. You send a prompt, the model reads all of it, picks the most likely next word, appends it, reads everything again, picks the next word, and so on. One token at a time, left to right, until it decides to stop.
The core mechanism that lets the model read everything at each step is called attention. For every token in the sequence, the model computes three vectors: a query, a key, and a value. You can think of these data structures as a bit more complex key-value stores. To generate the next token, the model compares the current query against every previous key, essentially asking “which past tokens are relevant right now?”, and uses the answer to weigh the corresponding values and build up context.
This is implemented (as you may all know by now) through the transformer architecture. Transformer layers are responsible for encoding the input sequences into a meaningful representation, applying the attention mechanism, and decoding into an output representation. All LLMs are architectural variations of this basic cell.
To get a sense of each of these variations I highly recommend Sebastian Raschka’s LLM Architecture gallery: from GPT-2 to DeepSeek and GLM.
The keys and values for every previous token are recomputed from scratch on every single pass through architecture. If your conversation is N tokens long and you’re generating token N+1, the model recalculates N sets of keys and values it already calculated on the previous step. This is slow and wasteful in terms of the resources.
The obvious fix to this is to cache them. The query, key and values are computed once per token and stored so they can be looked up in subsequent steps instead of being recalculated. This is the KV cache, a running store of QKV tokens from all previous tokens stored in GPU memory (so they are readily accessible when needed).
The problem is that the KV cache grows with every token. With short messages this is trivial as all tokens fit in memory, but a long conversation, or a full code base, involves hundreds of thousands of tokens. Each token has its own key and value vectors, across every attention layer in the model, each stored as a full-precision floating-point number (as long as there’s no quantisation involved). For a model like Llama 3.1 70B, the KV cache for a single long context can consume more GPU memory than the model weights themselves.
This is one of the key bottlenecks in production inference. Serve more users simultaneously? More KV cache. Support longer contexts? More KV cache. Run cheaper inference? Figure out what to do about the KV cache. We are trading the compute necessary to compute on-the-fly the QKV values, for increased memory requirements.
By using quantisation instead of storing each value at 32-bit or 16-bit precision, one can round it down to 4 bits or 3 bits (or even 2 bits, like Microsoft recently showed). Some accuracy is lost in the approximation, but if it is not significant for the user case, the trade-off is obviously worth it. The question is how to do this well. Standard quantisation techniques add 1-2 extra bits of overhead per value as metadata, which partially undermines the compression you’re trying to achieve. Getting to genuinely low bit-widths without that overhead, and without accuracy degradation, is the hard part. HuggingFace has a really nice page with an overview of quantisation and a list of methods
Enter TurboQuant
But things may be about to change. Google announced this week TurboQuant. TurboQuant (see paper) is a two-stage algorithm. The two stages have different jobs.
Stage 1: PolarQuant. This is the main compression step. We currently store vectors using Cartesian coordinates as distances of a base to the
▸ 展开全文
03-29 11:25 · AI,技术,HackerNews,AI
Show HN: Public transit systems as data – lines, stations, railcars, and history
Public Transit Systems
Comprehensive information about public transit systems worldwide. Explore stations, lines, railcars, and historical data through a terminal-inspired interface.
$ Global Network Statistics
Total Stations
1,570
Total Lines
89
Track Length
1,506mi
Systems
9
> Available Systems
9 systemsLight RailLink
Baltimore, MD
The Light RailLink is a light rail system operated by the Maryland Transit Administration serving the Baltimore metropolitan area. Running 30 miles from Hunt Valley in northern Baltimore County through downtown Baltimore to Cromwell Station/Glen Burnie and BWI Airport in Anne Arundel County, it is one of the oldest modern light rail systems in the eastern United States. The system uses former railroad rights-of-way for much of its route and runs as a street-running railway through downtown on Howard Street.
Stations
33
Lines
3
Track Length
30mi
Daily Riders
14,500
Metro SubwayLink
Baltimore, MD
The Metro SubwayLink is a heavy rail rapid transit line operated by the Maryland Transit Administration serving the Baltimore metropolitan area. Running 15.4 miles from Owings Mills in Baltimore County through downtown Baltimore to Johns Hopkins Hospital in East Baltimore, it is the only rapid transit line in Maryland. The system features both surface and deep-bore tunnel sections, with six underground stations in the downtown core.
Stations
14
Lines
1
Track Length
15mi
Daily Riders
15,200
BART
San Francisco Bay Area, CA
BART is a heavy rail rapid transit system serving the San Francisco Bay Area in California. It connects San Francisco, Oakland, Berkeley, and surrounding East Bay cities with lines extending to San Francisco International Airport and the Silicon Valley.
Stations
50
Lines
6
Track Length
131mi
Daily Riders
323,000
Beijing Metro
Beijing
The Beijing Subway is a rapid transit system serving Beijing, China. It is the world's busiest and one of the longest metro systems, having surpassed Shanghai in total length in 2023. Originally opened in 1969 as China's first subway, it has undergone massive expansion, particularly after Beijing won the 2008 Olympics bid.
Stations
527
Lines
28
Track Length
551mi
Daily Riders
10.5 million
CTA
Chicago, Illinois
The Chicago 'L' is the rapid transit system serving the city of Chicago and some of its surrounding suburbs. Operated by the CTA, it is the fourth-largest rapid transit system in the United States by ridership. The system's name comes from the elevated railways that make up much of the network, though portions also run in subway tunnels, at grade, and in highway medians.
Stations
146
Lines
8
Track Length
224mi
Daily Riders
594,000
NYC Subway
New York City
The New York City Subway is the largest rapid transit system in the world by number of stations, operating 472 stations across four boroughs of New York City. Opened in 1904 and serving approximately 3.6 million riders daily, it is one of the few transit systems in the world that operates 24 hours a day, 7 days a week, forming the backbone of transportation in the most densely populated major city in the United States.
Stations
472
Lines
25
Track Length
245mi
Daily Riders
3.6 million
Sound Transit
Seattle, WA
Sound Transit operates regional express bus, light rail, and commuter rail services in the central Puget Sound region of Washington state. The Link light rail system connects Seattle, Bellevue, and surrounding communities.
Stations
50
Lines
3
Track Length
59mi
Daily Riders
142,000
Tokyo Metro
Tokyo
Tokyo Metro is one of two rapid transit operators in Tokyo, Japan (the other being Toei Subway). Originally opened in 1927 as Asia's first subway, it was privatized in 2004 from the Teito Rapid Transit Authority (TRTA/Eidan). Tokyo Metro is the busiest metro system in the world by annual ridership.
Stations
180
Lines
9
Track Length
121mi
Daily Riders
6.84 million
WMATA
Washington, D.C.
The Washington Metro, operated by WMATA, is the rapid transit system serving the Washington metropolitan area. It is the second-busiest rapid transit system in the United States and serves the District of Columbia, along with portions of Maryland and Virginia.
Stations
98
Lines
6
Track Length
129mi
Daily Riders
626,000
Compare Systems
Side-by-side analysis of transit systems with detailed metrics and visualizations.
Open Compare Tool →
▸ 展开全文
03-29 11:25 · AI,技术,HackerNews,AI
Siclair Microvision (1977)
HomeSearchIndexesEquivalentsArticlesAdverts & BoxesTimelineCollectionsDonateContact
After years of rumour and many confident predictions that next year would see the launch, Clive Sinclair showed his 2 in television receiver to a large collection of professional cynics - the press - on Monday, January 10.
Let it be said at once that the Microvision is a remarkable piece of design. Most of the components, including three of the five integrated circuits and the tube, were developed for the receiver, which measures 6 × 4 × 1½ in. It is most certainly not a toy - with a £200 price tag it is certainly free of that connotation - since most of the design compromises appear to have been made.with low power consumption in mind.
Thetubewas developed by AEG Telefunken, who based their development on a design by A V de V Krause (Wireless World, July 1974, p.259). The screen diagonal is 2 in, and the tube length about 4.5 in. Electrostatic deflection is used to reduce power consumption. The EHT is about 2 kV, derived from an oscillator which provides barely acceptable brightness. (We were not able to look at the picture out of doors because the Sinclair staff seemed so nervous that some acquisitive scribe might make off with the set that we weren't allowed to hold it, let along go outside with it.) Sinclair were unsure of the heater power - at one time the Krause tube consumed 30 mW - but that of a similar AEG Telefunken tube advertised in German and Danish magazines is quoted as being 35mW.
Sinclairbs notion that the receiver could be of use to travelling business-men is supported by the flexibility of tuning and standards. The own-design bipolar tuners cover Bands I, III, IV and V with the aid of p-i-n diode switching and varicap-diode tuning. Line standards of 625 and 525 at field rates of 50 or 60 Hz are provided, with IF selection of 4.5, 5.5 or 6MHz. Tuner sensitivity, assisted by separate, folding VHF and UHF aerials, was good enough to give a reasonable picture in the steel-framed Savoy Hotel. Gated AGC and AFC are included, as is flywheel sync.
Sinclair decided to use two SGS i.c's for sound IF and 50 mW audio output, but vision functions are performed by three chips designed by Sinclair and a further 20 transistors. Attention to power-saving in the design has yielded a truly amazing consumption of 750 mW from 4.8 V. Four 1.2 V rechargeable cells are used. These last four hours but an external AC adaptor will charge them and run the set as well. A 40-hour battery pack can be used, as can a car battery, which will also charge the internal cells.
It may be said that the size of the picture is too small for comfortable viewing. It is little bigger than a 35 mm slide, which is usually projected or seen through a magnifier. Sinclair say that 2 in was selected after tests with many people and point out that 2 in at a viewing distance of a foot - the distance they say, at which most people read a paperback book - is equivalent to a 24 in screen 12 feet away, provided that the spot size is reduced to enable the resolution of the big picture to be retained, which they claim to have done.
However, the angle subtended at the eye is not the only important measurement. Normally two eyes will be used, and the way they converge, together with apparent size, is part of the process which gives the viewer an appreciation of distance. The brain will recognise a big picture, placed farther away than a small picture, as being bigger than the small one and will automatically assume that the bigger one is more detailed, even though it may not be. C Burns, in an article inWireless Worldin January 1953, pointed out that a magnifying lens not only increases the subtended angle but makes the screen look farther away, thereby adding to the apparent increase in size. Sinclair have probably considered the use of a lens, but if so they made no mention of one as an accessory.
The scepticism which has arisen as a result of the previous frequency of false alarms about the Sinclair set's imminent appearance in the shops was compounded by the secrecy which still surrounds its, production. At the Savoy press conference on January 10, Clive Sinclair said that they first started the project 14 years ago and had produced several complete receivers in that time but they all lacked something and we wanted to do it properly. We have spent £½ million of our own money to do this in research and development. Yet the National Enterprise Board, which last November took a 43% stake in Sinclair, worth an injection of another £650,000, on the basis of a three-year profit forecast on the Microvision, have not, Sinclair says, committed the company to any production schedule. He would not say when production would start, how many staff would be making them, or what the initial number made would be, yet he insisted that the sets would be in the shops 'next month'. Who would be stocking the set? He couldn't say: 'We haven't invited any orders because we have
▸ 展开全文
03-29 11:25 · AI,技术,HackerNews,大模型,AI
The Last Contract: William T. Vollmann's Battle to Publish an Epic (2025)
A few years ago, the novelist William T. Vollmann was diagnosed with colon cancer. The prognosis wasn’t great but he went ahead with the treatment. A length of intestine drawn out and snipped. It was awful but it worked. The cancer went into remission.
Then his daughter died.
Then he got dropped by his publisher.
Then he got hit by a car.
Then he got a pulmonary embolism.
But things are looking up.
William T. Vollmann spent “twelve or fifteen years” researching and writing a novel about the CIA called A Table for Fortune; as of this writing it has a few back-channel blurbs from editors and assistants who’ve caught glimpses and say it might be his masterpiece, or at the very least a new sort of achievement for him. But when he finished it, in 2022, he turned it over to his publisher, the final installment of a multi-book contract (although even that part gets complicated), and that’s when, to use Vollmann’s words, “Viking fired me.”
His publisher of thirty years.
It’s more complicated than that.
For starters, when he first turned it in, A Table for Fortune was 3,000 pages.
But word count wasn’t the only issue.
Vollmann’s daughter Lisa had a drinking problem. It was worsening for years. Homelessness, hospitalization, dangerous encounters. Vollmann, who’s never owned a cellphone or used the Internet, bought a burner phone so he could call her every day at noon. If she answered he would offer his studio, tell her she should sleep there instead of outside, on the street, or at the shelter where some woman had tried to kill her.
She’d tell him no.
Mostly, she didn’t answer.
Lisa died in 2022. A year later he wrote about the whole situation for Harper’s: about Lisa, Viking, the novel.
He concedes in the essay, regarding A Table for Fortune, that maybe the untamed sprawl and uncharacteristic number of typos were a sign that he wasn’t paying as much attention in the final stages as he normally might. Fine.
He’ll grant them that.
But there was other stuff too — and this’ll sound craven, to be so business-minded about books when really, as Vollmann will be the first to tell you, his work has never sold many copies anyway. His editors and copy editors and publicists are true believers. They love the work and believe in him. The word “genius” comes up a lot in conversation and not with qualifiers; just short declarative sentences, wincing into the first syllable. “Jean-yus.”
Plus they’d been together thirty years, him and Viking, and of course they were there through everything he’d just been through: with the cancer, with Lisa, his whole trek through this novel . . .
But Viking, in their defense, might lean on the case that, if writing novels is an art, making books is a business, and if Vollmann’s job is the latter then theirs is the former.
So for instance: A Table for Fortune employs a number of different fonts to signify different speakers, memoranda, newsprint.
Well, Viking doesn’t own those fonts. And they’re not free.
Vollmann is serious about turning his novels into beautiful books, almost tangible works of art, and if you fan through any of his recent tomes you’ll see it’s never static. Paragraphs go for pages but if you look at the sentences they’re these long serpentine things and then they’re not. They’re short. Clipped. The prose generally chatty and peppered with exclamation points and the typeface often changes and there’s some ALL CAPS writing and then a photo. A drawing. A chart.
Using the specific fonts that he wants in A Table for Fortune, as opposed to some of Viking’s in-house alternatives, would raise the price of each printed copy by “two cents,” as Vollmann claims in a recent appearance on the TrueAnon podcast.
Likely exaggerating.
He’s still prickly about the whole thing, quick though he is to say, in all seriousness, that he isn’t upset, it's fine, and to chant his new mantra, the one he mentions on TrueAnon and a couple other podcasts while promoting a book in 2023 that turned into a scandal.
It’s the mantra that, six weeks after that TrueAnon podcast, I’m disoriented to hear through my phone in a friendly, congested, one-minute voicemail:
“Nobody owes me a living.”
I spoke with Vollmann about A Table for Fortune last year, and when I told him I’d be editing the hour-plus recording down to a fifty-ish minute podcast, I asked if there were any parts he considered central, or anything he’d like me to omit.
“Just please don’t make me sound angry,” he said. About Viking, the fonts, the novel. “I’m not angry.”
He explained his remorse for the whole thing and a hope that they might work again some day if he’s still around. He said he can feel things “winding down.” With corporate US publishing, for sure, and maybe his career overall.
But anyway.
“Nobody owes me a living.”
So what’s this book about?
A Table for Fortune is a history of the United States Central Intelligence Agency. Half of it focuses on an intelligence analyst, Dave, and the other half on Matthew, his son.
“When Bill mentioned that h
▸ 展开全文