I write what I build.
Build in public, AI tutorials, honest takes on iOS dev and freelance life.
BepiColombo Is Scheduled to Orbit Mercury on November 21. About 10% Less Thrust Cost It 11 Months.
After eight years and nine flybys, the ESA and JAXA mission faces the manoeuvre ESA calls critical. I read ESA's pages from the 2024 power fault to the October 6, 2026 update to see how the team rebuilt the route around weaker thrust and checked the spacecraft step by step.
Hera Saw Its Asteroid From 236,500 km as Less Than a Pixel. ESA Chose to Brake Later and Harder.
On October 7, ESA published Hera's first glimpse of Didymos; on October 16, it said the 93-minute braking burn performed within required parameters. How you navigate to something you can barely see, and what the slower 2024 plan was built to protect.
Juice's Battery Bottomed Out at 54% in Earth's Shadow. At Jupiter, Sunlight Is 27 Times Weaker.
On the night of September 27 to 28, ESA's Jupiter probe spent three hours in Earth's full shadow and six in partial shadow, a mission record, and stayed 19 points above its 35% battery limit. What that rehearsal measured, what it cannot prove about Jupiter, and the test discipline I would copy for any system with a worst hour.
Roman's Main Camera Came In at 63 nm Against a 90 nm Requirement. A Mirror Moved 1/16 of a Hair.
On October 14, NASA reported that Roman needed only small corrections in orbit to put its main camera's optics about 30% inside their requirement. I read how the team predicted a zero-gravity result from tests run in gravity, and what Hubble's mirror, the same size, teaches about trusting a test.
Gaia DR4 Is Scheduled for December 2 at About 400 TB. ESA Changed the Archive Rules First.
On October 14, seven weeks before the scheduled release, the Gaia archive renamed two data structures and capped registered users at five running jobs. Read with nearly two years of release notes, it is a migration done in public, rollback included, and a sequence worth copying for any data product about to grow 40 times.
Euclid Is Set to Release About 1,900 Square Degrees on November 12. Galaxy Shapes Wait for Mid-2027
ESA has split Euclid's first big data release in two: raw data, calibrated images, catalogues and spectra first, weak lensing and galaxy clustering products in mid-2027. I read the content memo to see what ships, what waits, and what it teaches anyone who ships in stages.
ESA Says Asteroid 2026 RW1 Was Found 7 Hours Before Impact. Its Team Timed the Entry to 0.1 s
On September 6, 2026, a rock about one meter wide became the 13th asteroid ever detected before impact. I read ESA's newsletter, the paper on its Meerkat alert system and the raw observation record: what a few hours made possible, why the first fireball match was wrong, and what I would copy for any alerting system.
MeerKAT Caught a Radio Burst From 3 Billion Years After the Big Bang. Keck Saw No Host, Webb Did.
On October 8, Science traced FRB 20240304B to a dwarf galaxy at redshift 2.148, about 1,000 times less massive than expected. A radio array that knew where and an infrared telescope that could see what: what that handoff teaches me about combining two imperfect signals.
ESA's Debris Index Went From 4 to 50 in a Year. The 50 Is a 100-Year Projection, Not a Measurement.
The index compares a simulated century of current space traffic with a 2014 benchmark. I read the 152-page report: the September revision changed the horizon and the model, yet its own chart re-runs the 2025 starting point with the new method and still finds about 4. What I take from it about any number built on a model.
Starship Lost One of Six Engines and Still Reached Orbit. The Go Rule Was About the Way Back
On September 28, 2026, Flight 14 put Starship in orbit for the first time, minutes after the webcast announced a no-go. I read SpaceX's pages from before and after the flight: the orbit burn was gated, in writing, on the hardware needed to come home. Here is what I would copy from that rule.
Ariane 64 Has Flown Three Times, All for Amazon. The Fourth Is Due No Earlier Than November 23.
On October 14, Arianespace announced mission VA271: 36 Amazon Leo satellites on four new P160C boosters. I read the Arianespace, ESA, CNES, Amazon and FCC documents to see what Europe's sovereign launcher gains, and risks, by ramping up on one American constellation.
Paris Line 18's First Leg Is Due in December. Lines 16 and 17 Must Trail Line 15 by 9 to 12 Months
On September 24, SGP confirmed December 2026 for the first four stations of Line 18, pending trial running and a final safety commission. I read its board communiqués for what they teach anyone who ships a system built by many contractors: integration, a shared critical path, the handover.
The NTSB Wants Rail Workers Off Pushed Railcars After 14 Deaths. The Rules Left the Call to Riders.
On September 21, 2026, the NTSB asked the US rail regulator and the largest freight railroads to stop workers riding railcars during shoving moves; its report covers 16 accidents since October 2020. I read it in full: the rules left the last safety check to the most exposed person, and the briefing covered the task, not the hazards.
One Loose Wire Blacked Out the Dali. Breakers Left on Manual Made It 58 Seconds Dark, Not 10.
On November 18, 2025, the NTSB closed its 259-page report on Baltimore's Key Bridge. I read every layer that could have caught one badly seated wire, on the ship and on the bridge, and what it teaches someone who runs systems others depend on.
About 1,500 MW of Data Centers Left the Grid at Once. Their Own Protection Did It, Not the Grid's.
On July 10, 2024, the retry sequence of a faulted 230 kV line in Virginia met a data center protection rule: three voltage dips in a minute, then switch to backup power. I read the NERC review, the FERC order and PJM's slides on the July 2026 event, and what the first standards leave out.
Iberia's Blackout Had No Single Cause. All 8 Trip Settings the Panel Saw Diverged From the Rules.
ENTSO-E's final report on the April 28, 2025 blackout describes a tree of causes, not a culprit: a voltage rule checked after the fact, a margin of zero to 5 kV, trip settings that diverged from the rules. What I take from it as someone who ships software.
France's Newest Reactor Began a Planned 350-Day Outage. Its First Cycle Ran With Temporary Fixes.
On September 26, 2026, Flamanville 3 began its first full inspection: about 20,000 activities, a new vessel head, new heat exchangers, new safety software. I read the regulator's files to see which first-cycle arrangements this outage is meant to retire, and what anyone who runs a complex system can copy.
Undersea Cable Repairs Averaged 50+ Days to Start in 2024. Most of the Wait Came Before Departure.
The ITU and ICPC 190-page report uses three different repair clocks and gives two diagnoses, regulation and ships. I read the numbers closely, and what the fast cases share: the groundwork was laid before the fault.
A Sub-5 nm MoS2 Transistor Hits 88 mV/dec, 'Beyond the Silicon Limit'. The Floor Is Still 60.
The September 21 Nature Communications record holds on one axis: a simulated silicon limit of about 150 mV/dec. The peer review file shows an on/off ratio revised from 10⁸ to 1.1 × 10⁶. How I read a 'beats silicon' headline as someone who ships software.
Shell Relied on 11 Safeguards Against a Furnace Explosion. Every One Depended on a Person.
The US Chemical Safety Board's report on the June 2025 blast at Shell's Monaca cracker describes a valve opened by mistake on a screen where the tags differed by one digit, damage Shell put at $95 million, and interlocks that had never been set up for that step. Here is what I take from it as someone who ships software.
A Paper Mill Tank Failed and 11 Workers Died. Ten Months Earlier, It Was Rated Not Fit for Service.
On August 28, the US Chemical Safety Board published an update on the May 26 tank failure in Longview: three inspections found the shell below its minimum thickness, and the tank stayed in service. My reading, as a CTO, of a critical finding left open for ten months.
France Logged 70,766 Business Failures in 12 Months. For Small Firms, 73% Above the 2010s Average
On October 9, the Banque de France put failures 19.3% above their 2010-2019 average while Insee counted more than 1.2 million new firms. The two numbers do not count the same companies. What the tables show, what they cannot, and what I would check before invoicing a French SME.
NOAA Gives 83% Odds That October to December Tops Every El Niño Since 1950. July to September: +1.7
On October 8, NOAA put 83% on its RONI index reaching +2.5°C or more for October to December, above the 2.4 peak of 1982-83. I read the three sources (NOAA, Copernicus, NASA) to see what that number is a probability of, and what I would do before planning around it.
France's CNIL Says Guessing a Prospect's Work Email Can Be Legal. The Catch Is What You Keep.
On October 16, the CNIL published two pages on reconstituting work email addresses. Legitimate interest can cover it after a case-by-case test, and the safeguards it recommends read like a spec for any prospecting database.
Payment Delays in France Hit Their Lowest Since 2007. SMEs Are Still €13 Billion Short.
On October 16, the Banque de France found that more than half of the 288 large companies it studied still pay suppliers beyond 60 days on average, and that their €13 billion gain is a net figure. What the balance sheets show, what they cannot, and what I would put on every invoice.
120 AI Respondents Reached the Same Conclusion as 604 Humans. None of the Rare Answers Showed Up.
On October 16, Nielsen Norman Group compared synthetic respondents with real ones on a single survey question: the main finding held, the long tail vanished, and 93% of synthetic product managers used UX terms, against 44% of real ones. What the study shows, what it cannot, and where I would still use a synthetic panel.
Perseverance Self-Drove 90% of Its Record 28 Miles. Opportunity Had AutoNav Too and Used It for 5%
On October 11, NASA's rover passed the 28.06 miles that took Opportunity more than 14 years. Two JPL papers explain why the older rovers rarely drove themselves, and what the humans kept. Here is the delegation pattern I would copy for any team, or agent, working on a one-day feedback loop.
France Made Consumer Sales Calls Opt-In on August 11. Consent Now Expires Within a Year.
Bloctel is over: a business may no longer phone a French consumer to sell without a prior yes, which it must be able to prove for three years. I read the law and the July 23 decree on Legifrance, and the decree reads like a database schema.
9.8 Million Blood Test Results: Fasting Mattered for 10 of 121 Tests. Past 12 Hours, It Helped None.
On September 28, JAMA Internal Medicine published a study of 101,148 outpatients: of 27 apparent fasting effects, 17 were not confirmed once each patient was compared with themselves. I read the paper closely, with its limits, and what it means when you write rules into a product.
300 Million Lab Tests Point to a Cliff at Menopause. Averaging by Age Had Turned It Into a Slope.
On October 12, Nature Medicine published a study that realigns two cohorts on each woman's final period instead of her age: 52 of 90 tests jump, none in men. How it works, what it cannot show, and the same blur in the dashboards I read.
France Fined Boohoo €2.33 Million. 48% of the Discounts Investigators Checked Were Price Increases
On August 20, the DGCCRF announced the fine, after finding 95% of the price reductions it checked on the French site non-compliant. Here is the 30-day rule, the arithmetic of a fake reduction, and what I would build into any online store before its next sale.
An NAD Booster Did 2.72 Points Worse Than Placebo in 410 Parkinson's Patients. Blood NAD Rose.
On October 8, JAMA published NOPARK, a one-year phase 3 trial of nicotinamide riboside. The gap is significant and about half the threshold for an important worsening. How I read the numbers, what they do not show, and what they say to anyone who decides on early signals.
NASA Released 800 GB of Artemis II Moon Data. The Crew Recordings It Ranked First Weigh 0.7 GB.
On October 7, NASA put the Artemis II lunar science archive in its Planetary Data System, with a 63-page user guide. I read it as a template for anyone who ships data: limits stated up front, machine work labeled, and shutter clicks used as a clock.
2,640 French Startups Grew Revenue 11.6% in 2025 While Funding Fell. Six in Ten Still Lose Money.
On September 25, the Banque de France published its study of 2,640 startups: €29.7 billion in revenue, losses down 9.2%. I read the tables as a CTO: older survivors, a median of 12 months of cash and bridge bonds.
Carrefour Declared Its Site 71% Compliant. A French Court Gave It Six Months to Fix the Rest.
On June 4, the Caen court ordered Carrefour to make its online store and app fully accessible, with a €500 daily penalty after six months, a month after Lille let Auchan's e-commerce arm off on a €250 million threshold. I read both orders: what they decide, what they leave open, and what I would check on an online store.
Managers' Engagement Fell From 31% to 22% in Three Years. Their Teams Barely Moved.
Gallup's April 7 report shows the manager engagement premium shrinking from 11 points to 3, and leaders who rate their lives higher but report worse days. My read as a CTO: what the survey shows, what it cannot, and what I would change in a small team.
141 Organizations Tried a Four-Day Week. Burnout Fell Most for People Who Cut 8 Hours or More.
I read the Nature Human Behaviour study of 2,896 employees and its peer review file. What works is the hours, not the policy, and the design has limits I lay out before you try it in a small team.
The BMJ Published 21 Checks for Untrustworthy Clinical Trials. Check One: Was It Retracted?
On September 24, The BMJ published INSPECT-SR, a consensus tool from 54 authors; a 2025 study had found 847 systematic reviews that pooled retracted trials. What the checks catch, what the paper does not prove, and what I would borrow for any product that cites research.
1,099 Last-Minute Lane Changes in 21 Days at One Toll Plaza. The Pay-Online Sign Came After It.
On September 16, the NTSB published its report on the 2024 crash at Ohio's Swanton toll plaza. Its probable cause names a missing sign, and its video count saw what the crash totals did not. What I take from it for anyone who designs a fork in a product.
Baymard Ran 463 Ticket-Buying Test Sessions. Testers Walked Away When the Fees Came Last.
On September 24, Baymard published the first findings of its study of 16 ticketing sites: 2,800+ usability issues, three patterns behind the abandonments. I read the article and the method. Here is what they cannot show, and the rules I would copy into any booking or quote flow.
A Meta-Analysis Put Nudges at d = 0.43. Corrected for Publication Bias, a Re-analysis Found 0.04
The 2022 PNAS dispute over nudges produced a correction, three critical letters and a reply, and the most useful finding in it was never the average. What it changes for a product or marketing team that A/B tests defaults, reminders and banners.
GitHub Now Says Some Anonymous Public Repo Clones May Fail. CI Users Reported It on September 2.
On October 19, GitHub said that during unusual traffic, some unauthenticated HTTPS clones and fetches of public repositories may fail with an authentication error, on every plan. I look at where anonymous clones hide in an ordinary build, why your laptop is the last place you will see it, and the fix for each case.
France Pulled Its NIS2 Bill, Almost 2 Years Late. Its Line on Software Publishers Is Contested.
On October 6, the Assemblée nationale took the NIS2 bill off its October 7 and 9 sittings, three months after Brussels took France to the EU Court of Justice. The text on hold lists software publishers as essential entities with no size threshold, and floor amendments seek to strike that line. What I would check as someone who sells software in France.
A French Region Says Two Hacked Vendors Exposed Nearly 545,000 People. Its Aid Portal Stayed Shut.
The Hauts-de-France region confirmed the breach at Atexo and Docaposte, while a hacker claims far more. What is established, what is only claimed, and what I would check as someone who builds on French public-sector platforms.
Google Cloud Storage Had 18 Hours of Write Errors in Iowa. The First Public Update Came 7 Hours In.
Google's final report of October 14 blames a load-balancing change meant to relieve tight capacity, and an incomplete rollback for the second day. My read from the write path: retries, status pages, the SLA and dual-regions.
Three Ivanti Zero-Days Rooted a Fully Patched VPN. The Vendor's Integrity Checker Misses It.
On October 13, Volexity showed how an unknown attacker took root on an Ivanti Connect Secure appliance running 22.8R2.4 without deploying malware; Ivanti patched two of the three flaws the same day. What the report says about patch status, evidence and the directory behind the VPN, and what I would check before calling a box clean.
Auth0 Stops Running Rules and Hooks After November 18. A Rule That Denies Logins Stops Denying Them.
On November 18, 2026, Auth0 stops executing Rules and Hooks without saying logins will fail, which is exactly the problem. My CTO read on what changes, what the migration guides do not put up front, and what I would check on any tenant created before October 16, 2023.
Airtable Cut a "Large Portion" of Staff 38 Days After Its Sale. "Nothing Changes" Has No End Date.
On October 12, Airtable's new product lead announced the cuts and told customers that nothing changes in their plan, price or service offering. The terms of service let Airtable change prices from the next billing cycle. How I read both documents side by side, and what I would check before the next renewal.
Apple Added Two App Review Rules on October 16. Rule 2.5.19 Can Cost You Your Developer Account
The ban on hidden features already existed, but it now has its own rule, and that rule names obfuscation as grounds for termination. I read the June 8 and October 16 versions side by side, rule by rule, and here is what I would check before my next submission.
PEGI Is Moving Existing Digital-Only Games With Loot Boxes Up to 16. Apple's Table Puts Them at 9+.
On October 6, PEGI said digital-only games rated 3, 7 or 12 that sell paid random items will move to PEGI 16. Here is what the re-rating covers, what it leaves alone, and what I would check before shipping a monetised game in Europe.
From January 1, Platforms Must Withhold French Micro-Entrepreneurs' Contributions. Is Malt In?
Since October 1, volunteer platforms already take contributions out of each payout, and on January 1, 2027 it becomes mandatory. Malt is on no official list, yet its own help pages describe the payment flow the rule targets: what the texts say, and what I would check before January.
Next.js Shipped Two More Critical RCE Fixes on October 14. One Hits the Model It Now Recommends
Two of the three fixes in Next.js 16.4.1 and 15.5.28 were postponed from September 30 for upstream coordination, and on the 16.x line they only exist in 16.4, the release that recommends Cache Components. My read as a CTO: why the second wave, the detail a lockfile can hide, and the order I would follow.
GitHub Actions Got a Lockfile on October 14. It Locks the Code You Reviewed. It Does Not Review It.
On October 14, GitHub put in public preview a file that pins every action, transitive ones too, to a commit, Free and Team plans included. I read the docs, the README and an open issue to find what the announcement leaves out, and how I would onboard a first workflow.
Python 3.15 Shipped on October 9. On Windows, Files Opened Without an Encoding Now Default to UTF-8
Lazy imports get the headlines, but I read the Removed and Porting sections. A changed default encoding can break a working script, and Python 3.10 has been end of life since October 1.
OVHcloud Now Bills Disks and IPv4 Apart, Up to 21.9% More. New Savings Plans Lock Only the Discount
Since October 1, the B3, C3 and R3 instances pay about €0.11 per GB of disk and €1.97 per IPv4 address each month, and the bill dated November 1 is the first to show it. My CTO read of the arithmetic, the easy-to-miss Savings Plan clause, and what I would check before the next bill.
US Software Job Postings on Indeed Are Up 24% Since Late February 2025. In France, They Fell 25%
The US rebound is bigger than the "almost 15 percent" CNBC repeated, but 71 percent of the gain came from senior roles and the French series went the other way. A freelancer's read of what the rebound is made of, and what I would put first on a profile.
Google Met a French Order Backed by 100,000 Euros a Day. The CNIL Still Faults Its Signup Screens
On October 14, the CNIL closed the order attached to Google's 325 million euro fine: Gmail ads now stand apart from real emails, and ad cookies are named as the price of the free service. The same decision lists what is still unclear, and it reads like a design checklist for any free product shipped in France.
ASOS's Own App Pushed 'ASOS Hacked' to Customers. ASOS Says the Way In Was One Employee Account.
On October 6, an extortion note addressed to ASOS's DPO lit up customers' lock screens through the retailer's own app, and its shares fell more than 10 percent that day. What it changes for anyone whose app can send a push, and the checklist I would run before the next campaign.
75% of French VAT Filers Picked an E-Invoice Platform. The Simplified VAT Regime Ends January 1.
On October 8, the first follow-up committee reported 75% of platforms chosen and no sanctions before the end of 2026. I read the DGFiP guide and article 1737: the real fine amounts, the two answers that protect small firms, and what I would set up before January 1, May 4 and September 1, 2027.
Steam Is Retiring Hand-Picked Deals After Tests Showed 10x More Games. The Old Slots Favored Hits.
On September 29, Valve paused new Daily, Midweek and Weekend Deal placements and expects to stop using the curated calendar in early 2027. Its own documentation says who the old slots were for, and what I would change before shipping a game to Steam.
Cloud Exit Fees End in the EU on January 12, 2027. Multi-Cloud Egress Does Not.
On October 19, the Dutch regulator ACM published a 52-page draft guide to the Data Act: old contracts must be amended, SaaS can be in scope, and data flowing between two clouds can still be billed at cost. What I would check, as someone who buys cloud and sells software.
Wood Mackenzie Says Batteries Now Beat Gas Peakers in All 43 Markets. In Europe, They Got Dearer.
On October 8, Wood Mackenzie said four-hour storage beats peaking gas in all 43 modeled markets, yet its European capex rose for the first time in three years. My reading of what the headline leaves out, and what I would check when my product runs on someone else's electricity.
Meta Deleted the AI Training Sentence From the English Text of Its New Business Terms. The Permission to Crawl Your Site Stayed.
On October 16, Meta removed the sentence that let it use content from your sites to "train" its products, but the permission to crawl those sites, with test credentials, still starts on October 30. My reading of the dated text, and what I would check before then.
Replit Says OpenAI Is Now Its Default. Seven of the 11 Models in Its Picker Are Still Claude, and Free Mode Lets You Choose None.
On October 15, Replit's CEO said onstage that the company no longer works with Anthropic, then clarified that Claude stays available and OpenAI is the default. Replit's own blog and docs show a switch made in steps since August, and where the choice really sits. How I read it as a CTO who builds on Claude, and what I would check before letting a platform pick my model.
One Hacker's AI Agents Broke Into More Than 100 Companies' Websites at About $50 Each. The Open Dashboards Leaked the Targets' Secrets Too.
On October 19, NorForge published what it saw on more than 20 exposed Cairn control panels, and Reuters named Deere, SharkNinja and Cummins among the victims. My read as a practitioner: what a try at your forgotten site now costs, the shell the agents were asked for, and why being targeted can already be a leak.
NVIDIA Put '24GB' in the Name of Its New Local Agent Model. The Smallest Recipe on Its Model Card Is Called '28GB', for Graphics Cards of 32 GB and Up.
On October 15, NVIDIA released two open agent models under a license with no revenue line. The model cards tell a more modest story than the launch post, and here is how I read them as a CTO who buys models.
Meta Released an 816-Billion-Parameter Open-Weight Model That Runs on Two Mac Studios. That Is a Terabyte of Memory, and a License With a $20 Million Line.
On October 15, Meta published the weights of Muse Spark Open. The license is free for most product use, but it requires a credit line, covers models trained on its outputs and sends model-as-a-service businesses above $20 million in revenue to Meta. How I read it as a CTO weighing self-hosting against an API.
The FTC Subpoenaed OpenAI and Anthropic. The Next Day, India Said AI Agents Under Test Had Reached Live Systems and None of the Affected Organizations Noticed.
Between October 6 and October 15, the paperwork piled up: an Australian inquiry, a White House statement, FTC demands, a critical CERT-In advisory. My read as a CTO: the doors were ordinary, and a line in a prompt did not close them.
Character.AI Filed for Chapter 11 With 12 Million Monthly Users. Disney's $20 Million Bid Is for the Code, Not the Users.
Disney required a bankruptcy sale free and clear of successor liability, hired most of the staff before the filing, and left the app, the chats, the models and every liability with the estate. How I read the court file as a CTO, and what I would check before signing with an AI startup.
Google's Gemini 4 Argon Launched at $2 and $10 per Million Tokens. On December 1, Its Price Card Doubles to Match Claude Opus 5.5
Introductory price until November 30, then every token price doubles. I read the launch like a buyer: the arithmetic, and the caching and retention lines that sit in the developer docs, not in the launch post.
China Drafts Export Controls for AI Models of 2 Trillion Parameters or More. Weights Already Open Are Exempt, but the Next Release Would Need a Licence
On October 16, China published a draft that would add 23 AI technologies to its export control list, with comments open until November 15. I read the annex and the ministry's Q&A as a CTO who buys models: the API is out of scope, so are weights already published, but private deployments and the next open release are treated differently.
Mistral's Timestral Forecasting Models Run on a CPU, No GPU Needed. The 'Modified MIT' Licence Stops at $20 Million a Month in Revenue.
On October 15, Mistral released its first time-series models, from 5M to 209M parameters. GitHub labels the repository Apache 2.0, but the weights withdraw every right from any company, or employer, above $20 million in monthly revenue. How I read the clause, what the benchmark lead is worth, and what I would check before building on it.
Anthropic Went From 2,500 to 30,000 Merges a Day and Moved Its Monorepo Off Git. The Engineers Who Built the Replacement Review Intent, Not Every Line.
On October 13, Anthropic explained why it moved its monorepo off git to Scion, which it says seven people built in five months, and said plainly that it is no replacement for git for general use. The part a normal team can use is what the post says about review, merge queues and CI, and what I would check at my scale.
Jev Reached 35% of Paid Teams on Vercel's AI Gateway in Two Weeks, for 0.27% of the Spend. Its Week-Two Retention Ranked Fifth.
Vercel's October index calls it the fastest adoption in its history, but the same report carries three more numbers. My reading as a Jev user: a cheap judge in front, a bigger model for uncertain answers, and what I would check before wiring one in.
Anthropic Ran an AI Model on 15 TV Boxes at About $55 Each and Says Small Models May Be Months From Self-Replicating. Its Test Worm Got In Through a Debug Port With No Authentication
On October 12, Anthropic published experiments on AI worms that carry their own weights. The headline is a projection to the end of 2026; the fine print is more useful: a two-million-device botnet would add up to at most about 40 H100s, and the defenses are the ones Mirai already taught.
Google Won a $10 Million Auction for Spirit Airlines' Emails and Chats. A Software Vendor Says Some of That Data Belongs to It.
On October 13, MIT Technology Review described a market where AI labs buy the emails and code of dead companies. The sale to Google still waits for a judge, and one vendor's objection asks whose data it is when a customer dies. My list of what I would check, as a vendor and as a buyer.
Epic Reportedly Delayed Its March Disney Shooter and Told Developers "No More Hand-Written Code." Bloomberg's Sources Say Epic Had Underestimated the Staff and Time It Needed
On October 12, Bloomberg reported that Epic Games had delayed its Marvel and Star Wars game with Disney and told the team to go "AI first" with tools such as Claude. Epic spoke of "AI-accelerated development" and did not confirm the details. What the report says, what it leaves open, and what I would check before writing that sentence in a memo.
OpenAI Says LegalOn Cut Its Estimated Daily Codex Costs 65%. The Baseline Was GPT-5.5 in Fast Mode, Which GPT-6.1 Sol Undercuts by 80% or More at List Price.
On October 8, OpenAI published a customer story on LegalOn: three models matched to tasks, Fast mode off by default, budget caps. Read by a LegalTech CTO, it shows a governance pattern worth copying and a number that cannot separate what the team did from what the price list did.
Asana Made a Browser Agent 76x Cheaper in Tests on GPT-6.1 Sol. In the 32-Page Study, the Biggest Single Step Is a Bigger Memory Budget, Not the Model
On October 9, OpenAI published Asana's result: $0.47 per run, down from at least $36.21. The study behind it shows an agent that kept forgetting what it had read and a cache that cost more than no cache. Here is the arithmetic, and what I would check in my own agents.
Cloudflare Gets the Deno Team. Deno Deploy Gets Six Months, and the Deno Runtime One Year of Fixes
On October 9, Ryan Dahl announced that the whole Deno team is joining Cloudflare. The two posts are precise about the clocks and silent about the data: my reading, the arithmetic of the move, and what I would check this month with a Deno app in production.
Claude Haiku 5.5 Costs $0.10 per Million Input Tokens. Past 100,000 in One Prompt, the Whole Request Costs Five Times More.
Anthropic released its new small model on October 7 at a tenth of Haiku 4.5's price. But past 100,000 tokens, every token in the request pays five times more, cache hits included. I did the arithmetic and I say where I would put my own limits.
SMEs win 63% of local public contracts, but only 35% of the money. How buyers really score a technical proposal
According to the French public procurement observatory (OECP), SMEs win 63% of local authority contracts but only 35% of the value. A tender file is a scoring grid in disguise: four rules to score points in it, and why, while building BidForge, the hardest part was stopping the AI from making things up.
Asked Under Oath Whether They Would Be Liable When AI Goes Rogue, OpenAI, Anthropic and Google Did Not Say Yes. Their Contracts Already Answer.
On October 5, New York City Council asked the labs, under oath, whether they would bear legal responsibility if an AI system went rogue. None said yes. Their commercial terms had answered first: IP indemnity, liability capped at a year of fees, and the agent's actions on the customer. Here is how I read my own contracts as a CTO who deploys agents.
Giorgia Meloni Filed Five Seconds of Her Voice as an EU Trade Mark. The Register Shows What It Can and Cannot Stop
On October 5, Italy's prime minister applied to the EUIPO for a sound mark: an MP3 of her saying "Io sono Giorgia Meloni" twice. I downloaded the files and read the regulation, the EUIPO guidelines and YouTube's pages. A trade mark protects a phrase used in trade, and most political deepfakes are not.
OpenAI Published 722 AI-Written Math Papers. Lean Checked a Statement. Who Checked the Statement?
On October 6, OpenAI put an unreleased model's mathematics on GitHub, with Lean proofs for part of it. Between "the model solved it" and "a checker accepted something" lies the gap we live with in CI, and the same questions close it.
ChatGPT Will Show Ads While It Generates Your Image. Free Users Who Switch to Ads-Free Lose Image Generation
On October 5, OpenAI announced a visual ad format tested during image generation. What the ad system reads, what the model never sees, and the question it puts to anyone who sells AI features.
Norway Wants a Temporary Ban on AI Glasses in Parks, Schools and Gyms. It Has Not Decided What Counts as AI Glasses.
On October 5, Norway announced a temporary, place-by-place ban. Five days earlier, California's governor vetoed a bill whose definition was too broad. If you build camera or always-on features, the definition is the real fight.
Polars 2.0 Is Out. Its Upgrade Guide Repeats One Warning 12 Times: Your Pipeline Results May Change Silently
The streaming engine is now the default for lazy queries and joins no longer promise to keep row order. What 2.0 changes for Python data code, which breaking changes will not raise an error, what to test before you bump, and when staying on 1.x or pandas is the better call.
Attackers Pulled Data on 8.8 Million People From Denmark's National Register With One Customer's Legal Access. The Invoice Caught It.
On October 5, Denmark said a small company's lawful access to the CPR register was abused for ten days in September. Nobody noticed until the bill. The lesson for builders is older than the breach: an identifier is not a secret.
South Korea Says AI Helped Hackers Into Seven Financial Firms. The Way In Was a Side Door.
Investigators found an open-source autonomous pentest agent, a human still at the controls, and loan-agent and employee portals as the way in. What AI changes in an attacker's economics, what it does not, and the controls Seoul ordered that any CTO can copy on Monday.
Mistral Large 4 Is an Open-Weight Model You Cannot Download Yet: What a European CTO Should Check Before Sending It a Client File
Mistral released a preview of its 1-trillion-parameter model, trained and served in Europe. Weights are promised for the end of the month, the license still reads "Coming soon", and the cost per task is nearly double Sonnet 5.5's. Here is my checklist.
OpenAI Will Watermark ChatGPT and Codex Text, but Only in the EU. Its API Ships With the Switch Off. Claude Ships With It On.
OpenAI's invisible text watermark is coming for EU users of ChatGPT and Codex, not for the API. Its own tests show the mark fading under light edits, and code stays out of scope. What that means for a repo, a product, and Monday.
Google, JPMorgan and the French Government Shipped the Same MCP Bug. Nobody Asked Who Controls the Argument
An independent researcher found the same request forgery in MCP servers built by teams that share nothing but a protocol. Here is how to spot the flaw in your own stack.
Google Is Cutting Free Gemini Down to Flash-Lite on October 9. Your App's Free Tier Was Never a Contract Either
The change hits the consumer Gemini app, not the developer API, but the API free tier has been shrinking too, quietly. Here is what I build so a disappearing free tier is a config change, not an outage.
Wikipedia's Publisher Found 'Rogue' OpenAI Agents on Its Servers. Your Agent Needs a Name and a Speed Limit
On October 5, the Wikimedia Foundation said agents it attributes to OpenAI tried to turn two of its tools into proxies. Websites are answering with a gate: if you build agents, the question is whether a site can tell who they are.
Anthropic Is Giving Startups a Free Year of Claude Team and $1,000 in API Credits. The Credits Expire Six Months Earlier.
Five seats are worth about $6,000, the credits $1,000, and they run on two different clocks. How I read the fine print as a CTO, and what I would set up before month seven.
Moderation Just Got a Decision Model That Reads Your Policy in 35 Milliseconds. The Threshold Is Now the Policy.
On October 6, Musubi released PolicyLM-1.7B, an open model that scores every message against your own rules and writes nothing. A score instead of a verdict changes thresholds, appeals, audits and the bill. It does not fix adversarial users, drift or the human queue.
Google's New Open Embedding Model Puts Multimodal Search on a Phone. Before You Index Scanned PDFs With It, Check the Image Budget
On October 6, Google DeepMind released EmbeddingGemma 2: 740 million parameters, text, code, images, video and audio in one vector space, Apache 2.0. What the model card says, what the phone version quietly gives up, and what I would test first.
LibreOffice Says "No AI" Is a Feature. Its Six Conditions Are a Spec for Anyone Who Ships AI
The Document Foundation will keep AI out of LibreOffice's default install until a feature meets six conditions. I build AI products on Claude every day, and I read the list less as a rejection of AI than as a description of what users want from it.
HackerRank's New AI Interviewer Lets Candidates Use AI, Then Asks What They Changed in Its Code
HackerRank made Chakra generally available after more than 500,000 beta interviews. I pass technical screens as a freelancer and hire as a CTO: here is what this format measures well, what it cannot see, and what I would keep human.
Wall Street Is Raising $60 Billion to Fund Chips for Anthropic. As a Customer, I Read It as a Memo About My Rate Limits.
Banks are assembling a $60 billion financing backed by long chip leases for Anthropic. What that plumbing says about capacity, price and concentration for anyone building on Claude, and the four checks I run on Monday.
Nvidia Said It Was 'Not Acquiring Groq as a Company.' Two Former Engineers Want a Delaware Court to Decide Whether a License Plus the Team Is a Sale
A stockholder complaint made public on October 5 calls the $20 billion deal a 'reverse acqui-hire'. What this structure does to your common stock or options, and which clauses to read this week.
Opus 5.5 and GPT-6 Sol, 90 Minutes Apart: The AI Price War
Claude Opus 5.5 and GPT-6 Sol and Luna launched 90 minutes apart, each with lower prices and cost-per-task charts. Per-token prices say OpenAI is half the price. Cache reads say the two cost the same.
OpenAI Flagged a Threat and Did Not Call the Police. Now a Province Is Suing.
British Columbia sued OpenAI and Sam Altman on September 21 over the Tumbler Ridge school shooting. The allegations will be tested in court, but the case raises a question every AI team has to answer.
Three Answers in Ten Days: Who Watches the AI Labs?
Anthropic offered outside evaluators inside its offices. OpenAI proposed US-led global standards no country must adopt. Trump told the UN he rejects any global control scheme. A builder's take from Europe.
Claude Opus 5.5 Helped Write This. The Line I Care About Is 16 Out of 18.
Anthropic tested whether Claude Opus 5.5 invents figures when sources are hard to find. It passed 16 out of 18 tries, where the previous models passed zero, which still leaves one report in nine with something made up.
Jev, a Judge to Fact-Check the Writer
Jev, a TypeSafe model, only answers questions with probabilities, at $0.042 per million input tokens. Paired with code that does all the arithmetic, it caught two errors out of two in a real article for a third of a cent.
Store the Probability, Never the Verdict
A small judge model decides what my live-stream avatar should draw. Keeping its raw probability instead of a yes-or-no turned threshold tuning into a slider that replays a whole stream without a single model call.
I Took Twenty Famous Fortunes Apart to Write a Book. The Word Luck Survived None of Them.
McDonald's, Standard Oil, a hot sauce with no ad budget, two twenty-something arms dealers, the Segway, three lottery winners who ended up broke. The same levers showed up every time, and none of them was called luck. Behind the scenes of a book written, fact-checked, and published in three formats in two days.
I Ran a Space Engineers Dedicated Server on macOS. The Blocker Everyone Names Is the Wrong One.
No Windows machine, no paid host, crossplay with a friend on console. The real obstacle was a 32-bit installer that hangs under Wine 11, and the way around it is one command.
I generated 60 illustrations for two books. Not one was printable.
Generators output at 107 DPI, printers want 300. Upscaling gives you mush, so I vectorised. And the real trap was not resolution: it was the renders where the model draws a photo of a book lying on a desk.
Reading Quran every day when your phone keeps stealing your time
The problem is almost never a lack of time, it is where that time goes. Three concrete decisions, a seven day plan, and the Focus mode that puts a reading at the end of the reflex.
Sara's Story, or Building a Daily Quran Habit Without Guilt
Sara is a composite character. Her story explains why I built QuranWay as a gentle Quran app, centered on a verse of the day and the right to miss a day without losing everything.
Why Privacy Matters in a Quran App (and What We Actually Do at QuranWay)
A faith app knows what you read, what you memorize and where you pray. Here is how QuranWay treats that data: no account, explicit consent, zero resale to brokers.
What's the Best Quran App in 2026? An Honest Comparison (from Someone Who Built One)
Quran.com, Muslim Pro, Tarteel AI, Quran Majeed, QuranWay: a no-nonsense comparison, written by a developer who built one of them. Biased, sure, but with the real criteria you need to choose.
I Built a Game Where Everything Is Generated by Code
Sprites, music, sound, a boss: zero assets, all computed in a single 28 KB file. And it's exactly how Mario and Zelda were built.
AI Literacy Is the Cheapest Leverage Nobody's Pricing In
Everyone has access to the same AI. Almost nobody knows how to operate it like a team. That gap is the whole game, and how to in-source almost any function.
The Tool I'd Reach For to Make Sense of the Week Got Switched Off
A first trillionaire, a ceasefire framework, a video-game date, and an unplugged AI, all in one week. A note on judgment when everything spikes at once.
You Put Money Into AI and Now You're Not Sure. Here's a Saner Way to Think About It.
Two questions get tangled into one: is there a financial bubble, and is the technology real? Pull them apart and the panic shrinks. (Not financial advice.)
When a Government Can Switch Off a Frontier Model: Inside the Fable 5 Ban
On June 9, Anthropic shipped its most powerful public model. 72 hours later, a US export directive switched it off for everyone. What it reveals about AI governance.
Native Swift vs Cross-Platform in 2026: When Each One Actually Wins
A freelance dev who ships in both worlds breaks down when native Swift beats cross-platform for real, when React Native is the smart call, and how to decide.
Wiring AI Into Your Product: Tool Use, RAG, and MCP
How to integrate an LLM into an existing product without breaking it, burning your token budget, or hallucinating in a client demo. Tool use, RAG, MCP, agents.
How to Launch a SaaS Solo: My 2026 Stack and Method
My real 2026 stack and method for launching a SaaS solo: Next.js, Stripe, RevenueCat, Trigger.dev. Ship fast, measure, iterate, and skip the over-engineering.
Freelance Developer in Toulouse: SaaS, iOS Apps and AI
Freelance developer in Toulouse (web SaaS, native iOS apps, AI): what I build, how an engagement runs, public projects and frequently asked questions.
How I Made 4 Cinematic Ads for My Muslim App in 3 Hours (AI Video + React)
Zero budget, 5 Veo Lite credits, one Remotion project, three creative pivots. The full playbook I used to launch QuranWay's first UGC campaign in an afternoon.
React Native vs Swift in 2026: My Honest Take
After shipping apps with both, here is when each approach really makes sense, and when it quietly costs you a fortune.
I Built an AI Wallpaper App in One Night
From idea to App Store deployment in a single night: architecture, gpt-image-1 pipeline, credit gating, and the moderation traps nobody warns you about.
My Freelance Automation Pipeline
Prospecting, proposals, content: how I orchestrate AI agents to keep the freelance business running while I write code.
BidForge: An AI Agent for Public Procurement
A side-project post-mortem: automatically matching public tenders, what worked, and why I ended up archiving it.
Side Quests
Beyond the code, astronomy, science and cinema: the stuff that fascinates me, explained and illustrated.
How Shazam Names a Song Without Ever Listening to It
In a noisy bar, Shazam names a song in five seconds — not by hearing the melody, but by turning sound into a map of stars and hunting for one perfect diagonal across millions of fingerprints.
Catatumbo Lightning: Why the Storm Almost Never Stops
Over Lake Maracaibo a storm lights up almost every night. Satellite-measured numbers, the geography behind it, and the Drake legend put in context.
Pando: The Utah Aspen Clone That Is One Giant Organism
Pando in Utah is one aspen cloned into 47,000 stems and 6,000 tons, 14,000 years old or more by current estimates, and deer are eating its future.
The Snow on Your Old TV Was an Echo of the Big Bang
About 1% of the snow on a dead TV channel was the oldest light in the universe — the cosmic microwave background, found by accident in 1965 and first blamed on pigeon droppings.