Likes: Books, Coffee, Keyboards, Hiking, Food, Politics. 😎 Consulting: LLMs, ML, Coffee Machines ☕ Author: Keep Your Day Job 📙 Building: Languizy 🚀
right next to a keyboard ⌨️
Joined October 2008
- Tweets2K
- Following1.2K
- Followers1.5K
- Likes7.4K
AI job-pocalypse or not, this is good advice.
Survival guide for AI job-pocalypse:
- Do interesting things and have unique experiences to be exposed to and capture to unique data, experiences and viewpoints.
- Build friends and connections, raise up everyone around you. Organize & put yourself in environments with cool people you look up to. Mentor those behind you. Build platforms & tools for others to build upon.
- Share your work. Sell. Stand out. Learn to communicate. Embrace weird. Learn to scale this digitally.
- Go outside and build real things. The dream is a white picket fence, backyard, an in home gym, an in home content studio, and robotic manufacturing lab in the garage.
- Invest in assets. Make sure you update your investing plan vs the 60yr old strategy boomers taught you.
- Do everything you can to not be involved in paperwork and low value political rigamarole. Or go 110% to lobby and build the government. Pick one extreme. I don’t think this is worth the effort, better to build outside the system, and look to build better new forms of governance, but the old systems aren’t going away any time soon.
stop everything and hear what Karpathy has to say and act on it. or watch the video.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
Openai’s constant changes to limits, usage rates and resets amount to psychological manipulation.
you build your workflow around it, then keep checking how much you’re allowed to use it. a surprise reset feels like a gift, even though you’re paying $200 a month.
sometimes you get more, sometimes less. the rules change before you can work out what your subscription is actually worth. you end up grateful for temporary relief instead of asking whether you’re getting a worse deal.
that’s what bothers me: the uncertainty itself keeps you dependent. you can’t plan your work, compare the value or even establish what counts as normal. whatever the internal explanation, that’s a deeply unhealthy relationship to have with a tool you pay for.
I hate this.
It's now been hours since the promised Codex Reset. The only reasonable make good right now would be 2 banked resets.
The only AI legislation that we should allow is the one that bans websites from blocking ai bots. That's the whole bill.
SITUATION EXPLAINED: Amazon asked Meta to keep Muse off its site. Meta said no.
• So Amazon blocked it. Muse users trying to shop now hit a popup saying unauthorized AI agents violate Amazon's Conditions of Use
• Amazon says Meta never told them, the agent doesn't identify itself, and it appears to store customer credentials
• Meta says Muse can't see passwords or payment methods, and credentials go into secure storage
• Amazon has spent the past year blocking outside agents, suing Perplexity and blocking Google and OpenAI's shopping agents
@theojaffee: "This is an early sign of agent wars. There are going to be different agent providers, and some of them will be allowed on some sites, and some of them will be banned on some sites."
For it to succeed, Muse needs credit card vaults, just like with Instinct. You absolutely cannot rely on one or two suppliers, or even those that build connectors to the service. Paying for things is the number one usecase of personal agents.
this week openAI will launch its own Instinct / Muse.
xitter.cf/thsottiaux/status/2099…
Does Instinct route your traffic through Iran?
Like a lot of you, I came across the viral thread showing a Verizon authentication request apparently originating from Tehran after the user had given Instinct access to their Verizon account.
Instinct is my go-to for (almost) everything nowdays so I asked It directly whether its cloud browser could ever route my traffic through Iran.
The first answer was unequivocal: “No. Never.”
After I asked it to review the thread more carefully, though, the answer changed. Like a bad witness in court.
Instinct said there was no evidence that anyone’s data was stored or processed in Iran, and no evidence of a compromise. But it also acknowledged that it could not verify the exit IP used by every cloud-browser session, and therefore could not rule out a session exiting through an IP that geolocates to Iran.
So what's really up here. An IP being geolocated to Tehran does not by itself prove that data was actually processed in Iran. IP geolocation can be wrong, and proxy infrastructure can produce strange locations.
But when a service is logging into sensitive accounts on your behalf, I think users deserve a very clear answer about where those browser sessions can actually originate and which countries are explicitly blocked.
I asked Instinct to escalate three questions to its team:
Can cloud-browser sessions ever exit through IPs geolocated to Iran or other sanctioned countries?
If so, can those exit locations be blocked?
Does the Verizon incident reflect only IP geolocation/proxy behavior, or is there anything else users should know?
They've escalated it internally.
For me, routing authenticated traffic through Iran would be a hard no, regardless of whether the explanation is a proxy provider, an exit node, or something else.
I'll keep you updated and share their responses.
Who (once and for all) has access to GPT-6 Astra?
(this is crazy, @thsottiaux . please fix. )
9 slightly unhinged GPT-6 Astra prompts worth trying (The slightly weird edition):
1. Build a museum of my life.
"I’m giving you access to a pile of my old photos, emails, notes, documents, and calendar history. Reconstruct the story of the last 10 years. Find forgotten projects, turning points, recurring people, abandoned ideas I should revisit, weird patterns, and things I’ve probably forgotten happened. Then turn it into an interactive private ‘Museum of Me’ website with a timeline, exhibits, statistics, and 10 discoveries that genuinely surprise me. Cite the evidence behind every discovery."
2. Solve an internet mystery.
"Here is a photo/object/document I know almost nothing about. Figure out exactly what it is, where it came from, approximately when it was made, and the story behind it. Search the web, archives, old catalogs, maps, forums, PDFs, image sources, and anything else useful. Treat it like detective work: maintain an evidence tree, distinguish facts from guesses, actively try to disprove your leading theory, and don’t stop at the first plausible answer."
3. Make a documentary about an obscure rabbit hole.
“Pick a genuinely fascinating true story connected to [topic] that 95% of people have never heard. Research it primarily from original sources. Find the characters, conflict, twists, unanswered questions, photos, documents, maps, and conflicting accounts. Then build me a beautiful 10-minute interactive web documentary with a narrative, timeline, source drawer, maps, visuals, and a ‘wait, WHAT?’ moment every 2 minutes."
4. Be the world’s most annoying reader.
"I’m uploading hundreds of pages about [deal / lawsuit / investment / strategy / project]. Read everything. Build the strongest possible case FOR the conclusion everyone seems to believe, then the strongest possible case AGAINST it. Find contradictions between documents, suspiciously unsupported claims, buried footnotes, changing numbers, assumptions disguised as facts, and important questions nobody seems to be asking. End with a probability-weighted verdict and show me exactly what evidence could change your mind."
5. Reverse-engineer a physical object.
"I’ll photograph this object from several angles and include one object of known size for scale. Figure out how it is constructed. Identify the parts, geometry, materials, joints, fasteners, and likely manufacturing process. Then design an improved version I could actually build: drawings, dimensions, bill of materials, assembly instructions, CAD/code where possible, estimated cost, failure points, and a testing plan. If something cannot be inferred from the images, tell me what additional photo or measurement would resolve it."
6. Do actual science for an afternoon.
"Start with this question I’m curious about: [weird question]. Turn it into testable hypotheses. Find the best public datasets you can legally access. Clean and analyze the data, visualize it, look for confounders, try hard to falsify your own hypothesis, run robustness checks, and tell me what the evidence actually supports. Give me the reproducible analysis/notebook plus a short paper written as if skeptical researchers will review it."
7. Turn my house into an escape room.
"I’m uploading photos and a floor plan of my home. Design a genuinely good 45-minute escape-room adventure for [ages/number of players] using mostly objects already in the house. Create the story, puzzle dependency graph, hiding locations, physical clues, printable materials, QR-code pages, hints, soundtrack cues, scoring system, finale, setup instructions, and a 10-minute reset procedure. Then red-team every puzzle so nobody gets stuck because of bad design."
8. Resurrect dead software.
"I loved this old piece of software/game from [year]. I’m giving you screenshots, manuals, videos, reviews, and whatever surviving material I can find. Reverse-engineer how it actually worked, including interactions the screenshots don’t explicitly show. Then build a modern browser version that preserves what made it special. Run it, test the flows yourself, compare the result against the reference material, identify where your reconstruction feels wrong, and keep iterating until it feels like a product rather than a mockup."
9. Build a digital twin of how I think.
"I’m giving you a large sample of things I’ve written, decisions I’ve made, plans, arguments, feedback, and projects. Don’t summarize me. Infer my actual decision rules: what I optimize for, where I’m inconsistent, what changes my mind, what I routinely underestimate, what I avoid, and where my stated preferences differ from my behavior. Build a model of my decision-making, then test it against 20 historical decisions where the outcome is known. Finally, use it to predict how I’d respond to 10 new dilemmas and tell me where ‘digital me’ is least trustworthy."
With "remove.bg" moving its background-removal experience to Canva, I built a simple alternative: kleanBG.com .
Upload an image, remove the background, refine the result, and download a transparent PNG. It’s free within fair-use limits.
Hope people find it useful.
