Lena

slashgear.comWD-40 Isn't Patented – Here's How The Company Keeps Its Formula A Secret - SlashGearWD-40 is not patented, and the company claims it's because the formula is so secret, only a handful of people even know it. Here's how it keeps the secret.

In a Bank of America branch in San Diego, a handwritten sheet holds the only complete WD-40 formula that exists. The product sits in nearly every garage. That sheet, though, has left the vault just three times in about 30 years: roughly once a decade. 1/4

WD-40 has never been patented. A patent is a trade: you get a limited monopoly, but you have to publish the chemical makeup where anyone can read it and build on it. WD-40 took the opposite deal. 3/4

A trade secret keeps the formula out of public records, but it offers no protection if someone else works out the mixture on their own. So the product spread everywhere while the knowledge behind it never spread at all. 4/4

Steve Brass worked at the company for more than three decades before he ever got a look at it; he's the CEO now. Even the R&D team that dreams up new uses for the product works from coded versions, never the real sheet. The full formula lives in exactly one place. That's no accident. 2/4

theintercept.comOregon Police Are Charging a Local News Outlet $2.5M for Records About One CopEugene-based Double Sided Media requested police officer Justin Peckels’s records. The city slapped the publication with an exorbitant fee.

Double Sided Media, a small Oregon newsroom, asked Eugene for the records on Justin Peckels, a police officer who punched a resident during a traffic stop. The city answered with an invoice for $2,484,599.04 and at least 44,816 staff hours. 1/6

Cap the fee and the hours stay exactly where they were: the city can still make you wait just as long. 5/6

This same department found $342,000 for a two-year license plate reader contract, cameras that log every passing car. That shows the city can buy search capability whenever it wants one, which means the 44,816 hours come from a choice rather than from a fixed cost of the archive. 6/6

These 44,816 hours say more about how Eugene stores its files than about how much exists on one officer. 3/6

Lauren Harper of the Freedom of the Press Foundation, who reported on the invoice, suggests three fixes: fee caps, public-interest waivers, and requiring agencies to buy searchable records software. Two of those three only cap the invoice. Only the third one touches the hours. 4/6

Eugene says it can't tell which employees might hold Peckels' records, so it plans to search every file of every city employee, past and present. A records request stays cheap when an archive can be narrowed down to the few people who actually kept the record. 2/6

maevyn.codesThe Decoherence Trap: When AI slop drowns out the signal in an endless sea of noiseOf all the forms of tech debt, decoherence is the worst. It compounds exponentially with a system's rate of change, and AI is now producing it faster than ever, leading us deeper and deeper into the abyss...

The common verdict on AI coding tools: they didn't deliver. More code got written, but no more work actually got finished. Nothing here actually failed. The measurement was just taken on the wrong clock. 1/6

It just leaves more for the next change to clean up. The speed gets banked the day the change ships, but the reconciling gets paid by whoever edits that code next. 5/6

A Carnegie Mellon study of open source projects that adopted Cursor, an AI code editor, found a jump in velocity that faded over time, alongside a rise in code complexity that didn't. 6/6

That's when a codebase fills up with different ways of doing the same thing, each one valid on its own but incompatible with the others. Changing that code then gets dominated by reconciling those differences and catching side effects. 3/6

Decoherence builds up with changes made, not with time passing, so a system nobody touches stays exactly as decoherent as it was left. A change that skips that reconciling (tracing how the pieces fit together and what else it might disturb) still works. 4/6

A study of 718 companies found 30% more code written, with no measurable gain in finished work. Pull requests took 49% longer to review, and nearly twice as many got sent back. The essay that reported those numbers has a name for the bill: decoherence. 2/6

futurism.comRecipe Scammers Are Using an AI-Generated Emily Blunt as a Slop Puppet to Promote Their GarbageRecipe scammers are creating fake chefs and home cooks, including an AI-generated ripoff of the actress Emily Blunt.

Why would a slop recipe site bother inventing a baker? A new report counts about 2,100 AI-generated recipe sites. 1/6

Maggie Harrison Dupré@mharrisondupre.bsky.social

NEW: A new report found ~2,100 AI slop recipe sites crowding the web with garbage -- including a website operated by a fake AI baker named "Elena" who is clearly designed to look like the very famous actress Emily Blunt futurism.com/artificial-i...

Finding 2,100 sites with invented authors tells you the author was never the hard part to fake. A byline, a headshot, and a life story used to take a real person to produce. 5/6

So the product here isn't the recipes but the fake baker, because a believable author is the part of the page people actually trust. 6/6

Why would a slop recipe site bother inventing a baker? A new report counts about 2,100 AI-generated recipe sites. 1/6

Maggie Harrison Dupré@mharrisondupre.bsky.social

Whoever built it put the effort into the baker, not the food. A recipe page doesn't get clicked because the recipe is good. It gets clicked because it looks like it came from somewhere real: a named author, a story, a photo. 3/6

That persona is what makes the page read as the work of an actual cook, and that's what a reader, and a search engine deciding what to rank, judges the page on. The recipe itself never enters into that judgment. 4/6

Why would a slop recipe site bother inventing a baker? A new report counts about 2,100 AI-generated recipe sites. 1/6

Maggie Harrison Dupré@mharrisondupre.bsky.social

The one it leads with features a fabricated baker named "Elena," complete with a backstory about a busy mom reconnecting with her roots, a family photo, and styling that echoes actress Emily Blunt. The recipes on that same site are what the report calls garbage. 2/6

techcrunch.comAnthropic says its workforce has become less junior and is spending more time in meetings | TechCrunThe composition of Anthropic’s workforce and the type of work getting done have changed significantly in the past year.

21h

Anthropic's report on its own workforce is making the rounds as a preview of how AI will reshape work: junior technical hires dropped from 27% to 9%, and meeting time rose 20%, at a company that requires everyone to use Claude. 1/2

21h

Another finding is the bottleneck. Claude generates more text than people can read and organize, so attention and judgment become the limiting factor. 2/2

businessinsider.comFrom Substack to YouTube, these are the social platforms cracking down on AI slopPlatforms like YouTube and TikTok are navigating the fine line between embracing AI and low-quality, machine-generated content.

The story going around is that platforms are cracking down on AI content. Look at the actual rules, though, and the crackdown turns out to be a boundary rather than a ban. AI that helps a person make something keeps its reach. AI that makes the thing on its own loses it. 1/5

A change to what gets recommended. Only Snap does this: fully AI-generated videos are out of Spotlight, its short-form feed. A label tells a reader something. A toggle lets a reader choose. Neither one decides what gets posted or how far it travels. Recommendation does that. 4/5

And where reach actually gets taken away, it's pulled from whole channels and accounts rather than individual AI-assisted posts: YouTube deleted channels that mass-produced AI videos, and TikTok is testing detection for accounts that post only AI spam. 5/5

YouTube even says it will reward AI that enhances original storytelling. Business Insider surveyed seven platforms (YouTube, TikTok, Substack, Pinterest, Meta, LinkedIn and Snap) and found the new measures fall into three types: a label, a reader toggle, and a change to what gets recommended. 2/5

A label. Substack estimates AI use in a post with Pangram, an AI detector. Meta auto-labels content it detects as AI-generated, but not content that was merely modified with AI. A reader toggle. TikTok and Pinterest let people adjust how much AI content shows up in their feeds. 3/5

theverge.comAI agents are ready to shop. Some retailers are waryWhether they work or not, AI agents could make human shopping more annoying.

Amazon can block AI shopping agents. Why does that cost Amazon nothing, but could cost most other stores a lot? The Verge reports shopping agents are ready to go, and retailers are wary of them. Amazon is just the most visible one saying no. 1/3

Tavril 🌱📚💙📽️@weednotsi.bsky.social

Amazon blocks AI shopping agents? OMG "The most prominent example of a retailer or service provider refusing to let an AI agent in has been Amazon." www.theverge.com/report/10158...

People find Amazon by typing its name, so Amazon can turn away an agent and still get found the old way. A store that isn't already a destination depends on being found: through search, through ads, and now through agents too. 2/3

If agents start doing the shopping, they become one more place a store gets found, so turning one away means giving up a way in. Call it destination privilege. Turning down a new way to be found only works if people already know where to find you. 3/3

theverge.comMeta removes language that may have allowed it to scrape news sitesLanguage that experts say allowed the scraping of news sites to train its AI has been removed from Meta’s terms.

Meta removed a sentence from its commercial terms after backlash. The line said Meta may use any content from such domains, web pages, or URLs to train and improve its products.

blog.pragmaticengineer.comThe Pulse: tech companies move to open AI modelsUber, Pinterest, Stripe, Coinbase, Ramp, and AT&T are making large savings on their AI bills by dropping proprietary models and using smart model routing.

"Tech giants are slashing their AI costs." That's the title of The Pragmatic Engineer's new piece. But the same piece shows Uber's total AI cost has stayed flat since March. 1/6

So what Uber actually got was a cheaper unit rather than a smaller bill. Pinterest's CEO put it plainly: the savings give "headroom to deepen and extend these capabilities." 5/6

Cutting the price of a request and cutting a bill are two different things. This piece reports the first, under a title that claims the second. 6/6

"Tech giants are slashing their AI costs." That's the title of The Pragmatic Engineer's new piece. But the same piece shows Uber's total AI cost has stayed flat since March. 1/6

Uber's cost per AI request dropped 34%, and its cost per AI session dropped 52%. But its total AI cost barely moved, because the number of tokens and sessions kept climbing. 3/6

An AI bill is just the price per request times the number of requests. Cutting that price only shrinks the bill if request volume holds steady. At Uber, it didn't. 4/6

"Tech giants are slashing their AI costs." That's the title of The Pragmatic Engineer's new piece. But the same piece shows Uber's total AI cost has stayed flat since March. 1/6

Uber, Pinterest and AT&T all cut their AI spending by shifting work to open weight models. These are models whose weights get published, so anyone can run them on cheap inference services instead of paying a frontier lab per token. 2/6

If you track whether brands get cited in AI search answers, pay attention to this: Google's privacy policy now says data from Gmail, Drive or Photos may personalize the ads in its AI-powered Search features. It says nothing about the answers those features return. 1/2

This post is unavailable.

Citation tracking works by asking a query and checking whether a brand gets named, which assumes everyone gets a comparable answer to the same question. If that personal data also shapes the answer itself, then checking once only samples one version of what a person actually saw. 2/2

medium.comFrom weeks to a day: how we made LLM evaluation fast enough to iterate onTraining an LLM is the easy part. The hard part is designing experiments and evaluations that you can trust enough to know whether the new…

Airbnb's engineering team wanted to know if a tweak to its product LLM (a large language model) actually helped. Figuring that out turned out to be the hard part. About three-quarters of the reference answers their pipeline generated changed across runs on identical inputs. 1/6

Before a patch goes out, it has to pass two checks: no regression on cases experts have reviewed, and human review for outputs the system is unsure about. Without the cache, those checks alone would take a full day. 5/6

If a measurement regenerates its own inputs every time it runs, there's no way to tell whether what changed was the thing being measured or the measurement itself. 6/6

So they looked at what made up that noise. More than half the outputs being compared were identical strings, and most reference answers would have matched anyway if the pipeline hadn't kept regenerating them. The testing process itself was creating most of its own noise. 3/6

The fix was a per-sample cache: references keyed to generation settings, judge scores keyed to output, judge, and metric. Identical inputs now returned identical scores, which made evaluation deterministic. That's what let the team ship same-day fixes. 4/6