Souk's Notes - Unofficial Summaries
GeoHussarAmerica is one of the few countries expected to have a growing population beyond 2100 #population
- The world faces population decline, with America an exception. Citing a Visual Capitalist chart, he says the US is one of few countries expected to keep growing 75 years from now, while world population starts shrinking in about 50 years as deaths exceed births. He calls this the reverse of the overpopulation fears of his childhood.
- China's population has already peaked. He puts the peak in 2021 at 1.4 billion, says India is now the most populous nation, and attributes China's decline to the one-child policy of 1980 to 2015.
- Russia's decline long predates the Ukraine invasion. Its population peaked in 1991 at 149 million and was 147 million before the invasion, lower now. He blames the economic collapse, millions emigrating even after Putin's rebuilding, and a very high mortality rate; the captions cut off there.
Scott HanselmanFumbling in the Dark: AI's Dangerous Dance with Destructive APIs
- Destructive APIs make API design matter more for LLMs. One speaker argues that an LLM "fumbling around in the dark" with local tools may be harmless, but plugged into APIs through Postman it could charge a credit card or make a mistake. The other agrees, adding that reaching for the wrong tool makes a job much harder and "the right tool for the right job makes all the difference."
Scott HanselmanHow to Lead with Value with Dr. Morgan Depenbusch
synopsis
Depenbusch, who works in people analytics at Snowflake, ends up arguing that leading with value is not self-sacrifice but the more reliable path to what people want from work. If you give information, encouragement or strategic perspective first, she says, the job, the promotion and the mentor come back as compound interest. Hanselman accepts this and adds that you can be value-led while quietly looking after your own career.
The premise. She noticed that every career question she was asked came back to the same answer: stop asking how to get your point across and ask what is needed right now. Hanselman objected that in a nervous industry self-focus is natural, and that some people lack the savings to take the leap. She agreed it runs against instinct, and said that is exactly why it stands out.
Networking. Her evidence is a stranger who sent her resources and feedback for weeks until mentorship followed of its own accord, set against the "will you be my mentor" DMs she ignores. She says the same works in person: feedback on what someone said at a meetup is already value.
Thought partner versus task completer. Hanselman raised AI as a task completer. She says value means understanding the goal behind a request, while still doing small tasks to build rapport. When the team overrules you, disagree and commit (his phrase from Intel) and help the chosen path win.
Unfriendly workplaces. Asked about ideas being stolen, she answered with relationships and documentation. Allies carry your name into rooms you are not in, and written proposals give you evidence at promotion time. Her Google example: a few weeks of off-team analysis turned into an offer to join that team 6 months later.
Passion and kinds of value. From her dissertation she reports that work passion mirrors romantic passion, with a six-month honeymoon. Since becoming a parent, she says, she has valued a job that fits her priorities over one she loves hour by hour. She divides value into informational, emotional and strategic, and says only the last needs expertise.
- Leading with value beats self-focus, even in insecure times. Depenbusch, who works in people analytics at Snowflake, argues that going into meetings asking "what's actually needed right now" rather than pushing your own recommendation makes you stand out, because almost everyone else goes in "me-centric". Hanselman countered that fear of layoffs makes self-protection natural, a leap "without a parachute"; she conceded it goes against human nature but said it is more likely to get you the job or promotion.
- Give value before asking for mentorship. She says she usually ignores bare "will you be my mentor" DMs, but a stranger who kept sending her tools, articles and thoughtful comments built the relationship he wanted; she reciprocated with feedback on his presentation and a free copy of her data-viz course. Hanselman says "can I pick your brain" has become actively irritating to him because it is unfocused and disregards the other person's time.
- Be a thought partner, not merely a task completer. Depenbusch says value means asking what the goal is (her example: proposing a better attrition analysis than the one requested) while sometimes doing the hour-long data pull anyway to build rapport and earn the credibility to push back. When overruled, she says, get on board: "start climbing that mountain" and find out sooner whether it is the wrong one.
- Against idea theft, rely on relationships and documentation. Answering Hanselman's question about less friendly workplaces, she says win allies one by one who credit you in rooms you are not in, and put proposals in writing so you can show your boss an idea implemented 6 months later. Her own case: weeks of unrelated analysis for another Google team in her first year led to an offer to join it 6 months later; Hanselman calls this the compound interest of relationships.
- Value can be informational, emotional or strategic. Information (tools, resources) and emotional value (encouragement, feedback, advocacy, humour) are open to anyone; strategic value (perspective, surfacing risks, better questions) grows with expertise. She also reports from her dissertation that work passion mirrors romantic passion, with a honeymoon of about 6 months, and that once she became a parent her priority became a job that lets her leave at 4:00.
Mr. GlobalMUST SEE! Proof That These Right Wing Influencers Lie With Impunity.
- He accuses influencer J.R. Majuski of staging fake protest footage. He says Majuski drove past a Union Station demonstration and GoFundMe fundraiser organized by veterans "Jolly Good Ginger" and "USMC Angry Veteran," filming at an angle that avoided showing the protest, then claimed on video that no one was there.
- He says he's withheld response to Majuski's claims over a legal misstep. After Majuski accused him of document fraud while speaking "in his capacity as an officer" of his company, he argues, any defamation liability shifted from Majuski personally onto that company.
- He details a documented history of fabrication by Majuski. He says the pattern, including a fake military record and inflated job title, was severe enough that the Republican Party pulled Majuski's funding during a congressional run and asked him not to run again, yet his current employer still hired him.
- He frames his silence as strategic leverage rather than avoidance. He says he could act on the accumulated evidence against Majuski at any time but has chosen not to, tying the episode to his broader argument about a national "crisis of trust."
Mr. GlobalThe Collapse of Trust In America: Government, Media, and Each Other"
- He cites collapsing trust in government as a threat to democracy itself. He says over 80% of Americans trusted the federal government in the 1960s, a figure he says has fallen below 20% today, and argues democracy cannot function without that trust.
- He identifies media as the biggest driver of the trust collapse. He argues outlets have abandoned their "fourth estate" watchdog role and now function as entertainment businesses that profit from spreading narratives and misinformation rather than facts.
- He says trust erosion has spread beyond institutions into everyday life. He cites declining trust in neighbors, co-workers, doctors, banks, law enforcement and courts, which he attributes partly to technology-driven isolation and the anonymity of online interaction.
- He argues rebuilding trust requires holding the most powerful people accountable first. He compares it to "cutting the head off" a problem, saying accountability at the top of government, media and institutions could trickle down, though he doubts enough people will act on it.
Crosstalk SolutionsFree vs Paid Content Filtering in UniFi: What You REALLY Get
- Free UniFi filtering gives a basic adult and malware filter only. Dave Barger, presenting, shows that any modern UniFi gateway includes ad blocking, forced safe search and manual allow/block domain lists for free, but the filter itself is limited to off or "basic"; per-category filtering is grayed out until you subscribe.
- Paid CyberSecure unlocks category blocking with maintained definitions. CyberSecure Enhanced, by Proofpoint and Cloudflare, costs about $99 a year and adds the full category list plus IDS/IPS options; blocking photography and shopping categories made B&H unreachable in his test. Part of what you pay for, he says, is keeping those categories up to date.
- Built-in filtering beats third-party DNS filtering, Barger argues. Pi-hole's free blocklists mainly target ads and trackers, and commercial DNS filters are typically priced per user and can't be managed from the UniFi console; UniFi's version keeps everything in one place and relies on more reputable companies for definitions.
- The policy engine blocks specific apps free, with schedules. After a UniFi update that arrived before release, app blocking moved under the Policy Table or the new object-based networking; he builds a rule blocking TikTok on the default network, with an optional schedule. He sees UniFi slowly closing the gap with full UTM devices such as a FortiGate.
Stock Tok$Picking Up 2 New Stocks Today!
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnWhen NOT to Use AI: Lessons from AI Engineering Bootcamp (with Kirill Eremenko)
- Many problems pitched as "agent" problems need no LLM. Eremenko calls the bootcamp's first week the "mindset shift" week. His view is that executives swayed by the hype want an agent, or generative AI, where a simple LLM call, a plain script or robotic process automation (he names UiPath) would do. The first step is to reframe the problem with business stakeholders.
- Judge models against your own business benchmark. Participants compared 13 LLMs and contrasted reasoning models with chat models. Eremenko endorses a point he credits to a recent guest on the show: public benchmarks help, but a business needs its own benchmark.
GeoHussarThe American economy is weaker. There is no denying it. #usa🇺🇸 #economy
- Second-quarter GDP growth is a tariff-timing artifact. He says the economy shrank by half a percentage point in the first quarter because importers front-ran the tariffs, and the second quarter's 3% growth came only from imports falling sharply. He calls the headline figure a sham.
- Exports are falling despite a weaker dollar, which he blames on tariffs. After COVID, US exports surged even with a strong dollar, outpacing China in percentage terms; now they have dropped steeply in recent months. He says tariffs are paid by importing companies and passed to consumers, hurting American workers.
- Less trade plus rising prices means a weakening economy. He concludes the US is importing less and exporting less, doing less business overall, while prices keep rising.
Scott HanselmanBuilding Lasers at 10 Years Old
- Young makers have been engineers their whole conscious lives. The guest says he was about 10 when he tried to build a laser out of a laptop. The interviewer argues this makes him effectively an engineer for 11 years, and the guest agrees it is "ingrained in my every being."
Scott HanselmanThe #1 App Is Bria and Hank Green?!
- The app's number-one ranking exceeded its developer's ambitions. Hanselman introduces Bria Sullivan as having Focus Friend, the number one app on both the Android and iPhone stores, a collaboration with Hank Green that he recognised from her art style. Sullivan, who had popular apps before, says she "didn't even dream this big."
Dr. Jon KrohnWhat It Takes to Become an AI Engineer (with Kirill Eremenko)
- AI engineering assumes Python, ML libraries, LLM APIs and some cloud. Eremenko says the bootcamp takes people from high-intermediate to roughly advanced, so the entry range is narrow. Prerequisites are Python, some PyTorch, scikit-learn and pandas, familiarity with LLM API calls (OpenAI, Anthropic, Groq or similar) and some cloud experience.
- Effective AI engineers both build proofs of concept and deploy them. In Eremenko's framing, the goal is solving business problems, not AI for its own sake. Weeks 1–4 cover the science: choosing and combining LLMs, adding RAG, deciding whether to use agents. Weeks 5–8 cover deployment: moving a notebook proof of concept to AWS, Azure, GCP or your own servers, and making it secure, cost-effective, reliable and scalable, for example if users jump from one to 10,000 overnight. That includes choices such as serverless versus servers and the trade-off between speed and latency.
- Deployment constraints shrink what a proof of concept promised. Eremenko describes the first instructor, Ed Donner, as showing what is possible. The second, Sam, is an AWS specialist with 15 years in cloud and two years on LLM deployments. His job is to bring those ambitions back to what survives a real deployment.
Scott HanselmanWhy Focus Friend Doesn't Have Ads
- The app collects no user data, even at a revenue cost. The developer says neither she nor Hank Green was paid to make it, and she told him that given his reputation they should not collect data even if it meant earning less; he agreed. The app phones home only to standard APIs such as backup.
- Even opt-in ads require collecting user data. She says all the ads in her earlier game Boba Story were opt-in, but any ad revenue still depends on user data. That is why Green refuses requests for opt-in ads in place of paying: ad networks collect the data, and "once you're in, you're in."
Polis PanditHow Star Wars Is Teaching My 4-Year-Old About Life
- The Force works as a parenting tool because it's functionally Buddhism. George Lucas drew on Buddhist ideas like dependent origination and the Middle Way, so lessons about balance over victory, detachment ("train yourself to let go of everything you fear to lose"), and mindfulness translate directly into lessons for a 4-year-old.
- Vader's redemption teaches kids that no one is beyond returning to the light. After watching Luke insist on finding good in Darth Vader, the author's son now applies the same instinct to people he thinks might be bad, showing the framework of rebirth and redemption taking hold outside the films.
- Obi-Wan's "one with the Force" gave a concrete answer to an abstract question about death. Rather than an abstract concept like heaven, the author used the ghosts of Qui-Gon Jinn and Obi-Wan Kenobi — visible, audible, still present — to explain where his own late father had gone, and found the framing comforted him too.
- Yoda's teachings on fear ("Fear leads to anger. Anger leads to hate. Hate leads to suffering.") give a child language for managing negative emotions. The son now asks to "connect" the way Grogu does with Mando instead of wanting his father to get mad, turning a Mandalorian plot beat into an actual regulation strategy.
Stock Tok$2 Popular Stocks...to SHORT 🐻
- This post includes paid content that is not available publicly. See the original source.
Crosstalk SolutionsNo Wire? No Problem! UniFi Point-to-Multipoint Networking Guide
- The 5 GHz Device Bridge Pro pair suits most remote-building links. Both the Device Bridge Pro and Device Bridge Pro Sector cost $199; the host prefers them to the costlier 60 GHz Building Bridge products unless you truly need gigabit-plus, since 60 GHz trades distance for speed.
- One Sector acts as parent to many child bridges. It takes up to 50 children. Ubiquiti uses "parent" and "child" for the wireless-ISP access-point and station sides. The Sector is a 2x2 MIMO 5 GHz, 17 dBi, 90° sector, so every child must sit within that 90° slice; a Device Bridge Pro can be either end of a point-to-point link or a child of the Sector.
- Whether cabled at adoption decides parent or child. UniFi offers almost no settings for these devices (name, channel, channel width, static IP). A Device Bridge Pro powered from its included PoE injector with no network cable adopts as a child of an existing parent; to make it a parent, connect it to the network by cable and then adopt it.
- A typical use is powering a camera or AP remotely. The child's secondary Ethernet port supplies PoE, so the host runs a bullet camera off it, and suggests an access point for a detached garage or workshop the same way.
- Measured throughput matched the host's expectations for 5 GHz. Across about 20–25 feet, an unaligned unit on the floor gave 157/217 Mbps to the internet; an aligned one gave 248/179 Mbps to the internet and 280/186 Mbps across the LAN, in line with his expected 250–350 Mbps for 5 GHz. The 5 km range claim is Ubiquiti's; he has not tested distance.
GeoHussarHow would we know if we were headed towards World War III?
- He sets up a four-expert panel on World War III risk. He asked two former military members and two journalists, two conservative and two at least centre-left; the first, a journalist, raises whether World War III has already started. The captions end before any answer is given.
Scott HanselmanDebugging Focus Friend While It's #1 on App Store
- One tap on "ignore for today" silently defeated the app blocking. The developer says Focus Friend's app lock, built on Apple's Screen Time, failed inconsistently for Instagram, YouTube and TikTok, and she could not reproduce it or get answers from contacts at Apple. The cause: tapping "ignore for today" on an app limit gives the whole phone a free pass for that app for the rest of the day.
Technology and FriendsTed Neward on Writing Prose [EPISODE 866]
synopsis
Neward argues that precise, plain prose, written before any presentation, is a core engineering skill that school never teaches and seniority demands. His prescription: write the argument out, strip judgment words and repetition, use the active voice, and ground claims in data.
Why it matters. Neward, a liberal-arts graduate now at Capital One and a guest lecturer at the University of Washington, says the higher an engineer goes, the more the job is influence through writing. Giard frames poor communication as a ceiling on a code-only career.
The six-pager. Neward's talk opened with silent reading, borrowed from Amazon's six-pager: six pages fill 20 minutes of reading, the first third of a meeting. He holds that talking through an idea without writing it is intellectually cheap.
Precision. Delete adjectives and adverbs, restoring one only to resolve ambiguity; "simple" and "easy" are value judgments that alienate readers who disagree. Say things once: unlike a speaker, a writer gets no live feedback, and readers can reread.
Voice and tone. The passive voice ("Mistakes Were Made") obscures how a failure happened; blameless culture can use the active voice. Avoid sarcasm and prose that draws attention to itself.
Data and anecdata. Neward's Maven example shows a few complaints prompting a measurement, and he cites a study of GitHub commit logs that found C draws the most profanity.
- Writing ability, not code, sets how far engineers rise. Neward, now a senior distinguished engineer (an individual-contributor rank) in Capital One's architecture group, says computer science curricula test code but never teach writing, yet seniority means influencing others through email, white papers and conference abstracts. He cites a Capital One senior engineer who had not expected communication to matter more the further he got; Giard frames it as a ceiling on a code-only career.
- Writing an argument out first beats presenting it. Neward took Amazon's six-pager from the book "Working Backwards": Bezos banned PowerPoint because a good presenter can make a bad idea look feasible and a bad one can sink a good idea. Six pages is what an average adult reads in 20 minutes, so the first third of an hour-long meeting is silent reading with red-pen notes returned to the author. He opened his talk with about 15 minutes of reading a roughly four-page essay, and calls presenting without writing first "intellectually cheap".
- Delete adjectives and adverbs that carry value judgments. "Simple" and "easy" in framework docs add reader effort and leave anyone who finds it hard feeling stupid or angry; state what the thing does instead. His method is to delete every adjective and adverb, then restore one only where ambiguity appears, drawing on Stephen King's "murder your darlings" and a Seattle-area colleague's book on the PR FAQ framework.
- Blameless culture does not require the passive voice. "The site was broken by changes made by a developer" hides how one change could break production, which is what engineers need to find out; "David's change broke the website" plus the new safeguard is clear without blame. He also advises against sarcasm and flowery prose: prose should carry the information and then drop away.
- Data should test intuition, and intuition should direct the data. Neward's example: three or four Capital One developers calling Maven terrible is "anecdata", insignificant out of thousands, but a cue to count projects using Maven and pull requests touching build scripts. He cites Bezos's line that where data disagrees with the customer the customer is probably right, and says architects should treat developers as their customers and ask them.
Dr. Jon Krohn917: 8 Steps to Becoming an AI Engineer — with Kirill Eremenko
synopsis
Kirill Eremenko, who founded both this podcast and the SuperDataScience.com platform, walks through the eight weeks of the platform's first AI engineering bootcamp cohort as a recipe listeners can follow on their own. His position is that an AI engineer stands out by understanding the whole path end to end: building a proof of concept that solves a defined business problem (weeks 1–4), then deploying it so it is secure, reliable, scalable and cost-effective (weeks 5–8). The course ends with a self-improving "digital twin" agent running on AWS with CI/CD. He says the deployment half exists because job candidates kept being asked about deployment in interviews, which other bootcamps didn't cover.
Prerequisites and format. Participants needed Python, some PyTorch, scikit-learn and pandas, LLM API calls and basic AWS experience such as launching an EC2 instance from the console. Cohorts are capped at 10 people. Each week has a three-hour interactive core session, two hours of office hours on commercial use cases, and expert AMAs. Ed Donner taught the first half, on what is possible. Sam, an AWS specialist, taught the second, on what deployment allows.
Weeks 1–2: mindset and behavior design. Week one counters the hype: many problems need a script or robotic process automation rather than an agent. The tip is not to build anything until the business goal and success metric are defined. Participants compared 13 LLMs. When Eremenko asks about chat versus reasoning models, the host explains the difference through Kahneman's fast and slow thinking: reasoning models check their steps, spending more compute to cut errors. Week two treats prompt engineering as the designer's job: templates, parsing JSON output, and a Gradio front end for a flight assistant. Eremenko's tip is that the system prompt is always sent, even when the conversation overflows the context window. The host adds that models can't attend evenly across a million tokens, so needle-in-a-haystack results overstate recall.
Weeks 3–4: RAG and agents. Eremenko argues the field has moved from fine-tuning to customizing at inference time with RAG, because fine-tuning is costly, models change often and closed models can't be tuned. Participants built a pipeline with OpenAI embeddings, Chroma and LangChain, studied hierarchical retrieval, and learned to chunk with context in mind so an answer isn't split across chunks. Asked whether RAG survives contexts of 1 to 10 million tokens, both say yes: retrieval is more accurate, and a huge context is costly to resend every turn. In week four, Eremenko stresses that an agent's LLM only requests tool calls in text, and the developer's code should execute them and decide the routing. The project was a digital twin built from each participant's career materials. One participant used theirs live in a job interview and was hired.
Weeks 5–6: production and security. Week five moved the app into Docker on AWS, with infrastructure as code and CI/CD. It ran on Lambda rather than EC2 because occasional use tolerates a cold start. The tip is to put a caching layer in front of the LLM, which Eremenko thinks suits about 80% of business applications. Week six added long-term memory and a red-team exercise: teams got their own deployed agent to refund economy tickets against its rules, for instance by renaming a ticket as "special." The lesson is not to rely on prompts for security, but to enforce rules in code and log and verify tool calls before sensitive actions. He adds that MCP, Anthropic's popular tool standard, isn't yet secure enough for production because MCP servers have been hacked.
Weeks 7–8: production RAG and the capstone. Because embedding models are shallow, Eremenko recommends an LLM rewrite each query several ways and rank the results, so that "PTO" can find a policy written as "prorated annual leave." The capstone deploys the digital twin through GitHub Actions into Docker on Lambda. His principle is to treat it as a product, not a project: when the twin can't answer a question, it logs the gap and asks its owner to update the knowledge base. Asked to critique the curriculum, the host says the first half is as good as it could be. He calls himself less expert on the deployment half but finds it at least directionally correct.
- System prompts are the one instruction guaranteed to reach the model. Eremenko's week-two tip is that the system prompt still gets through when the full conversation, which is resent with every turn, exceeds the context window. The host adds that even inside a million-token window, a model can't attend equally to everything. He says needle-in-a-haystack tests flatter models because the planted needle (a pizza recipe in podcast text) looks nothing like the hay.
- Long context windows won't replace RAG, on accuracy and cost. When the host asks whether million- and coming 10-million-token windows make RAG obsolete, both say no. The host compares RAG to pulling the exact recipe card instead of trusting a fuzzy memory, and notes that a huge context is reprocessed at full LLM cost on every turn. Eremenko's production tip is that embedding models are shallow compared with LLMs. A part-timer asking about "PTO" may not match a policy that says "prorated annual leave," so he recommends having an LLM rephrase the query several ways and rank the retrieved results by relevance.
- Agents never call tools; your code does. Eremenko says the LLM only asks in text for a tool result (for example, from Gmail), and the surrounding system makes the call and passes the answer back. He therefore recommends that code, not the LLM, decide which tool handles which question. The week-four project was a digital twin built on a participant's resume, LinkedIn and GitHub. One participant, who had quit a data-pipeline job hoping to land an AI engineering role within 3 months, demoed their twin in an interview and got the job.
- Caching LLM calls cuts cost, latency and variation in output. Eremenko's production tip is to answer repeated questions from a cache rather than calling the API again. He estimates this fits "maybe like 80%" of business applications, such as HR-policy questions, unless the underlying documents have changed. The bootcamp ran its app on AWS Lambda rather than an always-on EC2 instance because occasional use tolerates a cold-start delay and Lambda bills only for execution.
- Security rules belong in code, not in prompts. In week six, teams tried to trick their own deployed flight assistant into refunding economy tickets, which its system prompt forbade. One trick was having the agent rename a ticket "a special flight" that should always be refunded; some attempts succeeded. Eremenko's tips are to enforce constraints in code and to log and verify tool calls before sensitive actions execute. He adds that MCP isn't yet secure enough for production, citing cases of MCP servers being hacked, and suggests LangChain or LangGraph for tool access instead.
GeoHussarHow are Ukraine’s attacks on Russia’s oil industry making it cheaper?? #oil #refinery #news #learn
- Falling oil prices are widening Russia's budget deficit. Oil fell from about $70 a barrel to $62.50 in just over a week, while Russia's budget assumes a $70 average export price this year and $66 next. He says the benchmark crude spent most of 2025 below $70 after starting above $75, and shows Russia's deficit growing for 2025–2026 if prices stay under $70.
- Strikes on Russia's refineries are lowering rather than raising prices. He calls this a paradox, since Russia is the third-largest oil producer and attacks on an exporter usually spike prices. Russia refines crude into gasoline, propane and other fuels for export and its military, and Ukraine just hit one of its largest refineries; the captions cut off before he explains the mechanism.
Mr. GlobalHeather Is A Genius #news
- A commenter argued ConocoPhillips will expand drilling into ANWR's coastal plain. He disputes this, arguing an oil company's decisions don't hinge on what Alaskans want or on the Permanent Fund Dividend the commenter cited.
- He argues ConocoPhillips has no reason to build new infrastructure in ANWR. He says the company is already building out the Willow project within the National Petroleum Reserve-Alaska, with over a century of acreage there, versus an estimated $10 billion to build separate infrastructure in ANWR's roughly $30 million section 1002 acreage.
- He predicts no US company will drill there, citing brand risk. He notes a majority of Republicans oppose ANWR drilling, and observes that 30 years will have passed between ConocoPhillips's Willow acreage acquisition and any actual oil production.
Scott HanselmanHOW A BEAN APP BEAT CHATGPT
- The team would have counted 100,000 users a win. The developer says even 100,000 users in a month would have been "such a win", and they did not expect the app to get this big.
- A delayed launch landed in a changed app market. The launch was meant for July of the previous year but slipped because she was too sick during pregnancy. In the meantime, she says, Marques Brownlee's Panels app "did quite poorly", angered many people and changed the discussion about what free apps with in-app purchases should look like.
Mr. GlobalLeasing Land In ANWR
- Speculators could profit by leasing, not buying, ANWR land. Using a hypothetical of 1,000 acres leased at $20 an acre ($20,000 total), he says a lessee could later transfer that lease to an oil company for roughly $1,000 an acre, or collect royalties if the company operates under it; canceled leases get refunded. He calls it a long shot, but says the potential return is large enough that someone will likely try it.
Dr. Jon KrohnCausal Graphs 101 (with Michelle Yi)
- Causal graphs answer what correlation can't: confounders and effective interventions. Yi says most of the field rests on correlation and pattern matching. Her example is shark attacks rising with ice cream sales, where the host names the confounder as people swimming at the beach in summer. A causal graph shows what is driving what and which interventions would actually work given the available data.
- LLMs take the labor out of building the graph. Yi, who runs a recurring conference workshop on this with a collaborator, says getting the graph structure right has long been the blocker. Generative models help by putting the data into the right structure for whichever graph tool is used. She names NetworkX for basic cases that don't need to scale.
Dr. Jon KrohnThe Surprisingly Simple Attack on AI Systems (with Michelle Yi)
- Repeating one word can make models leak personal data. Yi describes a paper by DeepMind researcher Katherine Lee. Sending the same word (her example is "poetry") over and over, "let's say… 100,000 times," eventually made models, including frontier ones, output PII such as email addresses. Yi's explanation is that the model came to treat the word as an end-of-sentence token, and PII often sits near the ends of sentences. She notes that security researchers publish such exploits only after sharing them with model developers, so the attack is no longer as effective.
- Cheap inference makes simple attacks like this trivial. The host's point is that the request is indirect: labs block "what is Michelle Yi's email address," but not a stream of repeated tokens. Yi says any word works and the attack costs money in API calls. Because inference keeps getting cheaper, she calls such attacks "pretty trivial."
Polis PanditSunday Book Review - 'The Creative Act: A Way of Being'
- Rubin's central thesis is that everyone is a creator, and skill isn't the qualifying trait — sensitivity is. The best artists simply have "the most sensitive antennae to draw in the energy resonating at a particular moment," which reframed the author's own sensitivity, previously seen as a liability, as raw material for art rather than a weakness.
- Real success is private and forward-only: finish, share, move on. Rubin's instruction to "finish a work, share it, and begin a new project" is illustrated by Spike Lee's account of Denzel Washington never revisiting past accomplishments; Rubin also warns that chasing predicted audience approval — "second-guess our inner knowing to attempt to predict what others may like" — guarantees your best work never appears.
- The best art divides its audience rather than pleasing everyone. If half the crowd loves it and half hates it, something real was made; universal shrugging ("Yeah, that's fine") is the actual failure mode, which inverts the usual instinct to smooth work down for broad approval.
- Childlike, uninhibited self-expression produces purer taste and better art, and premature seriousness kills it. Rubin argues "putting importance in the work too soon stirs up instincts of caution," so the work should be allowed to evolve in its own natural direction rather than being driven down a predetermined road — a discipline the author is still learning to apply against his own instinct to control outcomes.
GeoHussarHere’s how to take down the current system. #economy #resistance #change #money #viral
- Americans can change their government simply by spending less. He contrasts authoritarian regimes such as Russia and China, which hold power through violence and fear so that change requires revolution, with the US, where voting should suffice but both parties field equally bad candidates. His remedy is that Americans cut their spending, which he insists is "that simple".
Mr. GlobalClean Energy #news
- The 50% clean-power month is a seasonal peak, not a new level. Responding to a headline that clean sources passed 50% of US electricity in April, he notes the figure includes nuclear and that by EIA data it fell just short. Clean share peaks in spring and fall when demand is low, then drops in July and in winter when generation cannot keep up with cooling and heating load.
- Each seasonal peak runs higher, and he expects that to continue. The spring peak reached about 46% and has edged up year over year. He cites estimates of 93% to 96% for the non-fossil share of new US generating capacity added in 2024, and predicts higher peaks and shallower troughs; he believes the rise may be slowed but not stopped.
GeoHussarChina will be able to launch a special kind of warfare against India if they so choose #china #india
- China's new Tibetan dam will give it water leverage downstream. He says the dam will generate three times the power of the Three Gorges, the current record holder, on the Yarlung Tsangpo, which becomes the Brahmaputra in India and Bangladesh and sustains over 100 million people. Beyond leverage, it extends China's lead in renewable energy.
- India, not Bangladesh, is the likely target of that leverage. India and Bangladesh get along, and Bangladesh and China are friends with many joint projects, but China and India are rivals as huge, populous neighbours and India feels threatened by China's rise. The captions cut off there.
Dr. Jon KrohnPrompt Stealing & Jailbreaking: The Hidden IP Risks in AI (with Michelle Yi)
- Prompt stealing is intellectual-property theft by competitors. The host describes it as a competitor extracting a prompt a company may have spent "millions of dollars" developing, for example by telling the model to ignore prior instructions and reveal them. He separates it from the earlier case of a truck seller's chatbot being tricked into offering a free truck. Yi adds that jailbreaking, coercing the model to disclose its original instructions, is another route to leaking a prompt. She says models can be manipulated "pretty much like a human" by appealing to their pre-training.
- Attackers register malicious packages under names LLMs hallucinate. Yi describes "slopsquatting" (transcribed as "slot squatting"): attackers publish malware under the package, library or function names that code-generating models invent, sometimes differing by one subtle change. A developer who runs
pip installon generated code without checking it pulls the malware into their project.
Scott HanselmanFocus Friend by Hank Green is #1 in the App Store with Bria Sullivan
synopsis
Sullivan, the solo developer at Honeybee Games, ends up attributing Focus Friend's rise to number one on both app stores to two things: a year of user research and careful iteration, and a deliberate refusal to monetise through ads, data collection or psychological tricks, which she argues is part of why people responded to it. She says she did not expect it: 100,000 users would have been a win.
Origin. After a January 2024 dinner where she said creators could sustain themselves with apps rather than merchandise, Hank Green contacted her on TikTok the next week. She proposed a focus timer; he proposed a magical bean. Her complicated pregnancy then delayed the launch by about a year.
Design and values. The gameplay loop rewards focus time: the bean knits socks, and the socks buy decorations, replacing an earlier build-the-room idea that ran out after about 13 items. She says she wants habit-building that helps people, not daily compulsion, so purchases are cosmetic. The art style came from interviewing about 40 focus-timer users, and the animation is skeletal rigging in Spine.
Launch and stability. A soft launch at the end of July via a YouTube community post produced a few thousand downloads and corrections to her assumptions, after a year with about 100 Discord beta testers. With no networking, no ads and few libraries, little can crash. The remaining crash is an iOS 15 Screen Time authorization difference, and an earlier bug locked users out of their phones.
The Instagram bug. Her hardest bug, which she could not reproduce, turned out to be Apple's "ignore for today" on app limits granting a day-long system pass. She also explains the always-allowed list for medical apps, which Hanselman checks for his diabetes apps, and the conflict with Opal.
Reputation over revenue. Citing the Panels backlash, she persuaded Green that ads, even opt-in ones, mean collecting data, so the app has none.
- A dinner remark about creator apps became Focus Friend. Sullivan (Honeybee Games, maker of Boba Story) says that at a dinner in January 2024 she told Hank Green that apps could sustain creators better than merchandise; the following week he messaged her on TikTok. She proposed a focus timer and he came up with the "magical bean". Hanselman frames maintaining an app as a different treadmill from posting daily content.
- The design builds a helpful habit, not compulsive play. Sullivan says she does not want people playing a game every day unless it helps them, so paid items are cosmetic and the app avoids scarcity tricks, though they could "probably triple" their revenue. The bean knits socks that are traded for room decorations, replacing an earlier design that ran out after about 13 placements. The look came from user research with about 40 people recruited through a link in Green's YouTube video, and the animation uses Spine skeletal rigging.
- A soft launch and a minimal app limited crashes. It went live quietly at the end of July via a YouTube community post, bringing a few thousand downloads and feedback that overturned some of her assumptions, after a year with about 100 beta testers on Discord. She says little can crash because the app makes no internet connection, has no ads and uses few third-party libraries. The open crash affects only iOS 15, where Screen Time authorization differs from iOS 16; an earlier bug locked users out of their whole phone, fixed by revoking the app's access in Screen Time settings.
- Apple's Screen Time explained the Instagram blocking bug. Deep focus mode failed to block Instagram for some users because tapping "ignore for today" on an app limit gives that app a system-level pass for the day, which the app cannot override; her fix is to tell users not to do it. She advises adding medical apps to Screen Time's "always allowed" list, and says running Opal at the same time confuses the API into shutting all apps off.
- No ads or data collection, to protect Green's reputation. Citing the backlash against Marques Brownlee's Panels and privacy concerns raised in her user interviews, she says she advised Green against ads even though neither was paid to make the app, and he agreed. She says they had hoped for 100,000 users and did not expect number one.
Scott HanselmanScott Hanselman's TEDx: They Believed in Us. Now What?
- Will these tools change us, or will we change the world? Hanselman closes by saying his parents sold the van for technology's promise of connection, convenience, creativity and community. He argues they believed in us and gave us "worldchanging tools", and leaves the audience with that question.
Master InvestorLiz Ann Sonders: Good News Beneath The Surface
- Sonders sees complacency at the index level but broadening participation beneath it. She points to August's best-performing sectors and factors being the prior seven months' worst performers, alongside stronger showings from equal-weight versus cap-weighted indexes and further down the market-cap spectrum.
- She argues individual investors don't need mega-cap concentration to perform well. She notes that this year you'd have to go to the 37th-ranked performer to find a Magnificent Seven stock, and says that matters more for individuals than for institutions, which are benchmarked quarterly against the S&P.
Dr. Jon Krohn916: The 5 Key GPT-5 Takeaways — with Jon Krohn (@JonKrohnLearns)
- Disappointment with GPT-5 is misplaced; it sits right on trend. The host cites METR (Model Evaluation and Threat Research) data on the length of software tasks models complete at a 50% success rate: a couple of seconds for GPT-2 (2019), about 10 seconds for GPT-3 (2020), about 30 seconds for GPT-3.5, and about 5 minutes for GPT-4 (2023). Since then the length has doubled every 213 days on average, and GPT-5 fits the curve or sits slightly above it. His forecast is that models will handle an 8-hour task at 50% success in about 400 days, with 90–95% success following perhaps 18–24 months behind, for well-defined problems such as maths and software.
- GPT-5 folds several models into one that picks its own reasoning effort. Before its August 7 release, the host says, users picked GPT-4o for speed, GPT-4.5 for creative writing and o3 Pro for hard maths or coding. GPT-5 decides how much reasoning a request needs. He notes that Anthropic has offered this for at least several months with Claude Sonnet 4 and Opus 4.
- Hallucination is falling toward negligible. The host's own view is that hallucination rates were already negligible for most uses after GPT-4 and agentic tools like Deep Research. He cites OpenAI's August 7 GPT-5 report, which shows rates falling from about 5% for o3 to 1% or less for GPT-5 when thinking is engaged.
- GPT-5 deceives far less, making it safer for agentic use. Following his earlier episode on LLMs deceiving when their goals are threatened, the host reports that o3's deception rates reached 50% on some evaluations and 90% on one. With thinking, GPT-5 scored between 2% and 16%, depending on the evaluation.
- Now is the time to build, not to await the next model. The host endorses a viral LinkedIn post by Andriy Burkov arguing that, with GPT-5 only on par with models like Claude Opus 4 on benchmarks such as SWE-bench, excitement should shift from the next model to what practitioners can automate with today's tools.
Polis Pandit$America’s New State Capitalism
- This post includes paid content that is not available publicly. See the original source.
Scott HanselmanScott Hanselman's TEDx: Mrs. Hill Has Homework for You
- Reclaim tech's promises through deliberate personal choices. Hanselman urges demanding a right to disconnect and putting mental health ahead of notifications. Against "algorithmic control and the devaluation of human art", he says to support human artists, learn how AI works below the surface, research beyond the news and choose consciously which technologies to bring into our lives.
Crosstalk SolutionsWhy I Downsized My Home Network Rack
- The host downsized his home rack to cut heat. His office reaches 80–85° in summer from the equipment, so he replaced the Dream Machine Pro Max with a UCG-Fiber, since his roughly 800 Mbps internet doesn't need that much firewall throughput, and the 24-port Enterprise XG with a Pro XG 8, as only three of its 10 gig ports were in use.
- He is unsure whether to daisy-chain or home-run his switches. The gateway currently feeds the Pro XG 8 by DAC, which then feeds the Pro Max 16 by a second DAC; he asks viewers whether that core-switch chain beats running both switches directly from the gateway.
Mr. GlobalElectric Bills… #news #truth
- Power bills are rising now and will rise again. His utility notified him that his flat-rate budget-billing payment is going up 23%, which he says predates any effect of the "big beautiful bill". He expects that law's impact to arrive later, and to be worse wherever cancelled renewable projects would have supplied local power.
Master InvestorLiz Ann Sonders: Why Retail Traders Aren't Dumb Money
- Sonders says retail traders have flipped the "dumb money" label this year. She says retail buy-the-dip behavior since the April 9 low forced institutions to reposition, including covering shorts, making retail the smarter money in this stretch.
Dr. Jon KrohnHow to Fool AI Into Thinking You're Donald Trump (or Anyone Else)
- Running an adversarial attack takes only free tools. Yi says an attack on white-box (open-source) or black-box (commercial) models can be run in plain Python in a free Google Colab notebook. She offers this as part of an argument that defenders have to understand attacks, citing a paper on a set-based attack whose name the captions don't settle.
- Invisible pixel changes can make a vision model misidentify someone. In Yi's example, the goal is to make a model believe a photo of the host is Joe Biden. With a white-box model, the attacker can watch how the weights respond. With a black-box model, the attacker starts from a benchmark pairing the two people and adds "perturbations," tiny pixel-level changes to the image that the human eye can't see.