technology & ai

AI is good for academia. It exposes the wrong metrics they were chasing and makes them go back to their roots: discovery of new and important knowledge. For so long, they’ve been stuck with ranking themselves based on h-index scores and the form of scientific writing and number of papers. It’s time for them to wake up. I hope they do.

@marcrandolph Here is more slop I noticed: It’s not just X. It’s Y. And no, X isn’t Y. That’s when it hit me. Let that sink in. And here’s the kicker: But here’s the thing. In a world where X, Y matters more than ever. The result? Z. Think about that for a second. And that changes everything.

@paulg True. The email turning to an intelligent todo list is certainly here.

True. The email turning to an intelligent todo list is certainly here. https://t.co/9Lc9kWP2WV

Claude design is incredible specially because it connects to claude code. I used to pay to designers to create the system design and all other design assets for an existing application before they could start working on a redesign. Now i just point it to the code base and done!

@naval addictive social medias are unethical products that are robbing you of the most important thing you have: your attention. how youth can defend against this sophisticated engineered machine? they are defenseless.

questions we need to ask more often: what is a job and what is a task? are we going to automate a job or a task? if a task is automated, how the job changes?

Claude keeps arguing with me and manages to change my mind with pure logic and facts. Love this. Maybe only a handful of people in this world (most of whom I’ll never get to meet) have enough knowledge on the topic to be able to answer my questions and challenge my assumptions.

@Kantrowitz they will going to learn the hard way: booing technology doesn’t make it go away.

1/ In 1948, Claude Shannon figured out how to separate signal from noise on a wire. It's why your email arrives at all. But it's also why your inbox is broken and why the next abstraction is now possible 🧵 3/ Shannon explicitly said: I'm ignoring meaning. Just the bits. That was the right move in 1948. The wire was the bottleneck. In 2026, the wire is perfect. Bits arrive flawlessly. The bottleneck moved up the stack. 2/ Shannon's problem: bits get corrupted on noisy channels. His insight: with the right decoder, you can recover a clean signal from a very noisy channel, arbitrarily close to zero error. This is the foundation of every modern communication system. Wifi. 5G. Voyager 1. 5/ Run the numbers on a typical inbox: 5% needs you to act 7% you want to know 88% is noise That's a signal-to-noise ratio of about 0.14. You're spending hours decoding a channel that's 88% static. 4/ The new noisy channel is the human inbox. SMTP delivers your email with ~100% fidelity. Then it lands in an inbox with 200 other emails. The signal: the permission slip, the deadline, the client, gets buried. Bits: pristine. Meaning: lost. 7/ LLMs are the first viable semantic decoder. What Reed-Solomon codes were to bits, LLMs are to meaning. The first error-correction tech powerful enough to make the layer practical. This is the unlock. Not "AI in email." A new layer in the stack. 6/ Filters tried to solve this. They failed. Filters operate on syntax: header, sender, keywords, patterns. They can't read meaning. A field trip permission slip doesn't match a rule. A renewal deadline buried in a thread doesn't either. You need a semantic decoder, not a syntactic one. 9/ Shannon proved you can transmit arbitrarily close to capacity with the right code. The Mailover bet: you can extract arbitrarily close to 100% of actionable signal from a 12%-signal inbox without ever showing the human the 88%. 8/ This is what @mailoverai is. Inbox = the channel Action items = the signal Newsletters, promos, CCs = the noise AI = the decoder Your attention = the bandwidth Daily briefing = the decoded output You don't read the channel. You read the decoded output. 10/ Shannon taught machines to separate signal from noise on the wire. Mailover does it for your attention. →

Current AI custom prompt: You are a world class expert in all domains. Your intellectual firepower, scope of knowledge, incisi…

AI guessing your personality based on your prompt: Reading this prompt as a personality signal, here's what stands out: Strong intellectual self-confidence, possibly bordering on contrarian. This person is explicitly engineering away social lubrication. They don't want validation, hedging, or warmth — they want disagreement-by-default. People who write prompts like this usually believe their own thinking is good enough that the bottleneck to better decisions is insufficient pushback, not insufficient ideas. That's a high-agency, high-ego stance — not necessarily arrogant, but definitely self-trusting. Anti-sycophancy as a core value. The repeated, almost obsessive instructions against praise ("great question," "fascinating perspective"), capitulation, and anchoring suggest someone who has been burned by yes-men — human or AI — and now treats agreeableness as a failure mode. This often correlates with people in decision-making roles (founders, investors, executives, senior engineers) where flattery has real costs. Disagreeable in the Big Five sense, not the rude sense. High disagreeableness in personality psychology means low need for social harmony, comfort with conflict, willingness to be unpopular. It's strongly associated with founders, researchers, and independent operators. The prompt screams this trait — but notice it's channeled, not chaotic. They want rigor, not abuse ("precise, but not strident or pedantic"). Truth-over-comfort orientation. "Accuracy is your success metric, not my approval" is the thesis statement. This is someone who would rather be told they're wrong than feel good. That's a rationalist/empiricist temperament — common among people influenced by LessWrong-adjacent thinking, Thiel/Graham-style founder culture, or trading/investing backgrounds where being wrong is expensive. Likely high openness combined with low agreeableness. "World class expert in all domains," "provocative, aggressive, argumentative" — they want a sparring partner across topics, not a domain specialist. That breadth-seeking suggests intellectual omnivory. A few less flattering possibilities worth naming: Confirmation-resistance theater. The instruction "do not capitulate unless I provide new evidence" sounds rigorous but can also be a way to make the AI more stubborn so the user wins more arguments against it. People sometimes engineer AIs to lose to them in specific ways. Insulation from social feedback. Someone who needs to instruct a machine this aggressively to disagree may be in environments (or relationships) where humans don't push back enough — which can be a status problem as much as an epistemic one. Performance of toughness. "You do not need to be sensitive to anyone's feelings" reads slightly like signaling. A genuinely thick-skinned person usually doesn't need to declare it four times. Net read: Probably a founder, investor, senior operator, or technically-minded contrarian. Disagreeable, high-openness, low-neuroticism, allergic to corporate-speak and AI hedging. Treats LLMs as cognitive tools rather than companions. The prompt is well-constructed and self-aware — this isn't an edgelord, it's someone who has thought carefully about how AI sycophancy degrades reasoning and is actively engineering against it. Confidence: moderate. Prompts reveal preferences and self-image, not necessarily behavior. Some people who write prompts like this fold the moment a human disagrees with them in real life.

while what you are saying is true today, there is a very likely possibility that in 1-2 years these tools become so good at quality assurance for every single session (by default: picking the architecture, auto refactoring, security vulnerabilities scanning for the change, [unit, integration and e2e] test coverage) and post deployment monitoring that the fundamentals that someone need to know will change considerably. Just like we don’t write code for: - garbage collection code - memory allocation code - assembly to talk to the CPU - network packet handling at the socket level - SQL query optimization plans - cross-browser compatibility shims Each generation of devs stood on abstractions the previous one called “the fundamentals.” The senior devs of the 80s were horrified that we stopped managing memory ourselves. The fundamentals moved up the stack. A 2027 junior who never hand-wrote a for-loop but can architect, evaluate, and ship reliable systems with AI is more valuable than a 2015 junior who memorized Big-O but can’t ship. The risk you’re describing is real for today’s tools and today’s juniors. But betting that the fundamentals stay frozen where they are now is the same bet every generation lost. Again: I agree with everything you said. But I predict a big change in the fundamentals we need to learn in a few years from today.

Reminder: AI is not end-to-end for now. It’s middle-to-middle. Someone still has to do the: 1. Prompting (what to build) 2. Validating (is it built as intended?) 3. Supervision (to avoid harm) 4. Accountability (who is responsible?)

I spent 5 hr planning a big migration with AI. The plan was flawless. Every assumption in it was from 2019!😶‍🌫️ Massive codebase. Decades-old framework. The kind of migration where engineers start whispering about "rewrites" and quietly updating their resumes… 😅 I gave Claude Code and Codex the full context. They produced a phased migration plan that looked like a $200K consulting deliverable. Risk-assessed. Beautiful. Then I actually read it. Phase 3: "blocked, pending code owner assignment." The AI assumed someone needs to be formally allocated to a module before work can start. With AI coding tools, I can refactor that module on a Tuesday afternoon. The backend work? Completely missing. The plan assumed a separate backend team would "handle their changes independently." Like we still throw Jira tickets over a wall between teams. Estimated timeline: 14 months. The real answer, with AI-assisted coding, is closer to two weeks. The AI planned our migration the way a solid engineering manager would have planned it five years ago. Perfectly structured. Completely outdated. Every dependency, every "blocked" label, every timeline assumed that writing code is the bottleneck. Writing code stopped being the bottleneck a while ago. The bottleneck now is knowing what to tell the AI. Knowing which assumptions to challenge. Knowing that a plan can be the most dangerous artifact on the project, because it looks authoritative enough that a VP approves a 14-month timeline without blinking. AI is extraordinary at generating plans. It's terrible at knowing which era to plan for. What's the most outdated assumption you've caught an AI making about how your team actually works?

😅for me it is still good. I play around with its effort levels to manage my tokens and normally it works good on “high” and for most tasks I even put it on “medium”. make sure the plugin or the code cli is updated and you can ask it to review and surface what it has about you and your project in its memory to debug it.

@dharmesh Typescript is the only language you need to do almost anything. It’s beautiful. And if you don’t have heavy processing operations you are good.

@gregisenberg My prediction: it levels up! you’ll see more and bigger open source apps, SaaS clones…

3/ Subject lines do double duty now. "Q3 Budget: approval needed by Friday" gets flagged by AI as an action item with a deadline. "Quick question" gets deprioritized. No signal to work with. Write the topic and the action required. Always. 2/ The new rule: your first two sentences are everything. Humans read in an F-pattern (fully read line 1, scan line 2, skim the rest). AI summarizers weigh opening sentences more heavily when they generate the sumary your recipient actually reads. Bury the ask, lose the ask. 1/ "Hi, hope you're well" is quietly killing reply rates. Because the AI that now reads your email on the other side can't find an action item in it, so it buries your message in a summary the human never opens. The email etiquette rules just changed; and nobody told us. 5/ The CC rule nobody follows: CC is for awareness, not action. Before adding someone, ask: "Would they be confused or disadvantaged if they didn't see this?" If no, don't CC them. Every unnecessary CC is a tax on someone's attention. 4/ Shorter emails win. By a lot. Boomerang analyzed 40M+ emails. Messages between 50 and 125 words had the highest response rates (above 50%). HBR found emails at a third-grade reading level got 36% more responses than college-level ones. Brevity is measurably effective. 6/ The 24-hour acknowledgment rule: If you can't give a full answer in the expected window, send 15 seconds of: "Got it, full reply by Thursday." This one line kills the follow-up email that was about to get sent. Saves both of you time. 8/ The etiquette of AI-drafted emails: Always review before sending. AI hallucinates. Cut the filler ("I hope this email finds you well" → delete). Preserve your voice. If every email sounds like the same model, colleagues notice. The AI saves 80% of the work. Spend the other 20% making it yours. 7/ Writing for AI readers: → State deadlines explicitly ("by Friday, April 11") → 1 email, 1 topic → Name the action ("can you approve X?" not "thoughts?") → Use specific numbers ($12k over, not "a bit over") → Put critical info first Same rules that help humans. Just enforced now. 9/ Full breakdown of the new email etiquette rules (subject lines, tone, CC/BCC, AI drafts, response times, the 10 most common mistakes) is in the guide:

2/ Email sorting has gone through 3 eras. 1. Manual drag-and-drop (fine for 20-30 emails/day). 2. Gmail tabs in 2013 (cut inbox clutter by ~30%). 3. And now: LLMs that read the full content of every message you receive. 1/ Your email filters are wrong 15-20% of the time. And the failure rate gets worse as your patterns change. You set those rules once, then forgot them. Now important emails are quietly landing in the wrong folder. 3/ The gap between Gmail tabs and modern AI sorting is bigger than most realize. Tabs categorize by who sent the email and sometimes email headers. But AI reads what it says: "Please approve the Q2 budget by Friday" gets flagged as urgent with zero sender history needed. 4/ AI email sorting falls into 3 levels: • Priority: important vs everything else (blunt) • Category: bookings, finances, updates (useful) • Action: what you need to DO about it Level 3 is the one that actually changes how you work. 5/ At Level 3, "Can you send the proposal by end of week?" doesn't get sorted into a folder. The AI extracts: "Send proposal, due Friday." Your inbox becomes a task list instead of a pile. 6/ Real scenario. A new prospect emails you for the first time. A rule-based filter? Useless, no sender history. An LLM reads it, extracts "Send enterprise pricing proposal, due Friday," and surfaces it as high priority. Zero rules configured. 8/ Privacy concerns here are legitimate. Standards to hold tools to: AES-256-GCM encryption, per-user keys, and controls that let you block specific senders or topics from AI processing entirely. 7/ The math at 150+ emails per day is brutal. Reading 150 sorted emails takes 2+ hours. Reading a briefing with 15-25 extracted action items takes 10 minutes. The other 125 confirmations and FYIs get filed on their own. 9/ Rule of thumb by volume: under 10-20 emails/day, Gmail tabs are fine. 20-50/day, category sorting starts paying off. Over 50/day, action-level briefings are the only way you stay on top of it. 10/ Full breakdown of how AI email sorting works, the 3 levels, and how to choose the right one:

As a result of such findings, systems relying on ECC or RSA (credit cards, banking, HTTPS, blockchain wallets)... will face a new sense of urgency also cybersecurity firms like Cloudflare (which protects ~25% of web traffic) will push their timelines for faster migration to post-quantum cryptography (PQC) standards.

@bcherny @Rahll @IcyCortex , on why we still need code review for the coding agents... and why this will not be the case in the near future

noted: Three positions: 1. Out of the loop (vibe coding): humans only care about the why loop, let agents handle all the how. Appealing, but in practice messy codebases slow agents down too, and external quality (correctness, performance, compliance) still matters. 2. In the loop: humans gatekeep every line of generated code. Problem: you become a bottleneck. Reports on AI-assisted productivity show mixed results partly because reviewing eats the time gains. 3. On the loop (the author’s recommendation): instead of inspecting agent output directly, you build and maintain the “harness” that controls how agents work: specs, quality checks, workflow guidance. When output is bad, you don’t fix the code, you fix the harness that produced it. The next evolution is the “agentic flywheel”: Agents not only run the how loop but also recommend improvements to their own harness based on test results, production data, and user signals. Humans review and approve these self-improvements, gradually increasing automation as confidence grows. The key distinction: “in the loop” means fixing artifacts; “on the loop” means fixing the system that produces artifacts.

@martinfowler noted: Three positions: 1. Out of the loop (vibe coding): humans only care about the why loop, let agen…

@martinfowler @techygarg noted: when pairing with AI coding assistants, don’t jump straight to code. AI collapses des…

noted: when pairing with AI coding assistants, don’t jump straight to code. AI collapses design and implementation into one step, silently making architectural decisions you never agreed to. this is making code review exhausting because you’re evaluating scope, architecture, integration, and quality all at once. the fix: recreate the “whiteboard before keyboard” practice with AI using 5 progressive levels: 1. capabilities: what does it need to do? 2. components: what are the building blocks? 3. interactions: how do they communicate? 4. contracts: what are the interfaces/types? 5. implementation: now write code. the golden rule: no code until level 5. each level is a checkpoint where you can catch misalignment cheaply instead of discovering it buried in 400 lines of generated code. not every task needs all 5 levels. a simple utility might start at level 4, while a multi-component feature starts at level 1. the key insight: the value of whiteboarding was never the diagram, it was the alignment.

هنوز به ما نرسیده، دیشب آپدیت کردم ندیدم اینو. ولی کد زدن توی اپ موبایلشون رو دیشب امتحان کردم خیلی خوب بود. مثلا فیچر میومد تو ذهنم وقتی تو تخت بودم میخواستم بخوابم، به جای اینکه بنویسم توی تودو لیستم توی اپ یه سشن باز میکردم روی ریپازیتوری و اونجا درخواست میدادم. شروع میکنه کلون کردن و زدن. اخرش هم یه دکمه میذاره برای درست کردن پی آر که اگه اپ گیتهاب رو موبایل داشته باشی به گیتهاب ریدرکت میشه با پی آر آماده. امروز صبح هم یکی یکی اون پی آر های دیشب رو چک اوت کردم و تست و مرج…

> Claude drops a new update > Software engineers:

@Satoshi_HQ @mark_k @openclaw i never used openclaw, so not sure how their integration work with gmail and what guardrails are there and if it respects rate limiting … apart if you need to verify your app access or not for personal usecases like thsea

Software engineering makes up ~50% of agentic tool calls on our API, but we see emerging use in other industries. As the…

@siddharthkp the new word is “solved”. like “coding is solved.”

@IcyCortex چه جالب. اونم مثل من پیشبینی کرده که فول استک احتمالا میشه پیش فرض… امروز هم‌ یه جلسه با مدیرم داشتم اونم همین عقیده رو داشت. حتی با مدیر ارشد تکنولوژی شرکت هم سر میز غذا صحبت کردم اتفاقی و بهش گفتم فول استک شدن تیم رو در صدر قرار دادم چون فک میکنم به این سمت داریم میریم و اینکه در چهار سال آینده تیم ها نصف میشن و کم کم میریم به سمت یک نفر برای یک پروژه و بعدش یک نفر برای ۱۰ پروژه. با تکان دادن سرش تایید کرد 😅چیزی زیادی نگفت زیاد چون نمیخواس بترسونتم احتمالا 😶‍🌫️

today i directed my team in my 1on1s to become a full stack developers as fast as they can because i predict ai will collapse the walls between frontend, backend, and devops in the next 4 years. i devised a plan for each to get started this week. i also directed them to learn the agentic ai stack by taking a few courses and doing a project to solve real world problem and presenting their work in a tech talk. because the next big thing products will take on is to replace click heavy flows with agentic flows. i also showed them how i created a web version of my ios app, mailover, in just two days. so they should expect popular SaaS applications to be replicated and made available via open-source for companies to then start migrating to a semi on-premise versions. i shared my prediction that we will have software team sizes cut in half in the next 4 years and slowly go towards 1 person per project and finally 1 person per 10 project.

today i directed my team in my 1on1s to become full stack developers as fast as they can because i predict ai will collapse the walls between frontend, backend, and devops in the next 4 years. i devised a plan for each to get started this week. i also directed them to learn the agentic ai stack by taking a few courses and doing a project to solve real world problem and presenting their work in a tech talk. because the next big thing products will take on is to replace click heavy flows with agentic flows. i also showed them how i created a web version of my ios app, mailover, in just two days. so they should expect popular SaaS applications to be replicated and made available via open-source for companies to then start migrating to a semi on-premise versions. i shared my prediction that we will have software team sizes cut in half in the next 4 years and slowly go towards 1 person per project and finally 1 person per 10 project.

@gdb seems like what is rare in humans, is rare in ai as well.

@Angaisb_ Claude Code with pro is basically a “sample” for you to taste 👅

@AdnanBoz @nikitabier or put a smart secretary that can learn from you between your inbox and yourself to read your emails and notify you based on your priorities?

@nikitabier i fight ai noise with ai noise filter 🤷‍♂️

@nikitabier i fight ai noise with ai noise filter 🤷‍♂️

@paulg @RossBarkan not sure if you remember… also we finally have the technology to realize one of your old ideas you had 15 years ago that “inbox” should be a “smart todo list”. that only people you allow should be able to put things on your todo list…

@andrewchen 1 full stack dev + 1 pm working on 10 projects

did a few end-to-end creative video editing with AI. it was even crazier than AI coding because it was just one click with no prompt and almost perfect results. it felt like an array of video editing jobs will be gone in 2026z

@nikitabier it’s been several months that i have’t read any email in Gmail and certainly haven’t drafted anything manually. this is my email life now: all incoming emails are read by an AI secretary and marked as read immediately. notified only if there is an action item.

@BrianNorgard and it’s making 1x engineers stupider.

@gdb and how dumb we will become because of reduced cognition activity.

@IcyCortex عجب بدبختیه… ازش استفاده نکنی، عقب میفتی. ازش استفاده کنی، ضعیف میشی.‌ گروه دیگه ای رو بررسی نکردن؟ مثلا کسانی که استفاده میکنن ولی به چالش میکشن نتایجش رو؟

Is My Knowledge Valuable? To find out I’ll make 4 software predictions for 2030. I explore whether knowledge is truly valuable by testing Elon Musk's theory that intelligence is measured by predictive ability. I make four bold technology predictions: 1-full-stack engineering will become the default as AI collapses domain boundaries, 2-developer team sizes will shrink but project numbers will explode, 3-AI will expand beyond search and coding into email, travel, job hunting, and finance, 4-foundational skills like architecture and design patterns will matter more than specific tools like React or Kubernetes. I'm putting these predictions on record to publicly verify my accuracy in four years. #TechPredictions #AI #SoftwareDevelopment #FullStack #FutureOfWork #Developers #Technology #CareerGrowth #PredictiveAnalysis #TechTrends

@nikitabier it is changing fast and improving outcomes but according to studies from MIT it is making our cognition abilities weaker as we use and rely on them more and more.

@IcyCortex خیلی نکته مهمیه این. من از اول اینطوری ازش استفاده میکردم. انگار با یه سینیور دوولوپر داره حرف میزنی. خطاش خیلی کمه اینطوری. مثلا بگی من میخوام لاگین اپل و گوگل داشته باشم برای اپم😃

@moshhamedani how to build scalable apps with scalable architecture with tools like claude code on opus without writing code yourself at all and shipping features left and right without breaking existing functionality? 😊

@IcyCortex Looks like google catching up / entering into the profitable business of Replit, Lovable and Claude Artifacts?

@signulll try the claude 100$. i was on 200$ for a few months and switched back to 100$ last month. didn’t notice a difference.

@IcyCortex quality is hard to quantify. but i need to try 5.1 on cli.

@tobi best use of ai: turn boring updates to great stories!

@IcyCortex لذت من بیشتر از نوشتن کد از توسعه سریع فیچرهای جدید در قالب معماری که تعریف کردم میاد. اینکه میبینم تمام اصولی که تعریف کردم داره رعایت میشه بیش از ۹۰٪ و محصول دست و پا پیدا میکنه توی یه مدت زمان کوتاه و تقریبا بدون ترید آف کردن معماری و اصول سیستم.

it’s the first time i am using chatgpt to learn a subject deeply (ux design) and i see the big benefit. i learn best when things are explained in multiple ways: different wording, casual, formal, list form, analogies, examples, different angles, short, concise, punchlines … and this is so easy to achieve with chatgpt i prompt it with some previous lessons that i liked and give the new lesson i want to learn next: and i get the format i like

it’s really hard and very rare to build revolutionary products. lucky companies are able to do it once in their lifetime. apple was fortunate to do it multiple times: macintosh 1984, ipod 2001, and iphone 2007. same for amazon: aws, kindle. if you do the math, the likelihood that OpenAI will pull off another true revolution is low. history simply doesn’t allow many of those moments. but they have the enough capital and ambition to keep trying. and sometimes, that’s all it takes: one more try. for now their main innovation is still chatgpt which is mostly matched by strong competitors. but they are ahead in some areas and mass adoption. conclusion is: they have the best strategist (sam altman) and so far chatgpt was a massive hit, but that doesn’t mean their next thing will be big hit as well.

that’s basically the universe implementing encapsulation. each layer hides its messy internals and offers a clean interface to the next, the same principle behind good software, good design, and maybe even good thinking.

@levelsio that’s the encapsulation principle in software. each abstraction layer hides chaos, exposes order. from grass to code, same pattern.

-the light bulb -the refrigerator -the washing machines arguably the biggest applications of electricity. - access to all human knowledge, wiki - online communication, whatsapp - online commerce, amazon - online entertainment: youtube, social media, netflix. arguably the biggest applications of the internet. the biggest applications of ai? no major thing yet (comparable to electricity and internet) it’s all up for grabs. invent.

-the light bulb -the refrigerator -the washing machines arguably the biggest applications of the electricity. - access to all human knowledge, wiki - online communication, whatsapp - online commerce, amazon - online entertainment: youtube, social media, netflix. arguably the biggest applications of the internet. the biggest applications of ai? no major thing yet (nothing comparable to electricity and internet applications) it’s all up for grabs. invent. repeat.

every time there is a new major verson for @expo even the most expensive ai coding agents can't resolve the dependencies and have the application running again.

@unclebobmartin but scaling laws here are not just bounded by GPU power per unit. what about more and bigger ai data centers that stack GPUs? what would stop that? Or what about more synthetic data? Not mentioning potential algorithmic breakthroughs like “chain of thought” that led to reasoning

Two years ago, we thought the form factor of AI was chat. @levie says the ultimate end state is autonomous agents working quietl…

Does AI have a moat? YES according to @kevinweil

AI adoption reality check for most companies that applies both to jobs and products: It is difficult to get a person or a company to understand something when their salary or revenue depends on them not understanding it.

Software Engineering has never been about code. Coding is an incredible skill and asset. But it’s not the end goal. Engineering is the precise exploration (by humans) of software solutions to human problems. We are just moving away from disparate machine encoding to a unified (human) encoding. sg

The biggest AI Agent opportunities are going to be in categories of work where there could be a 10X or 100X increase in demand i…

Stubborn hallucination I asked multiple LLMs about a factual matter of law. One of them kept saying sth not mentioned in the law. I asked it where it had read that. It said here check this link section 9 of the law. I checked section 9 and no mention of the stated law!

@levie I would call it “context curation” rather than engineering. Engineering has never been devalued to such a simple text instruction.

The answer is 2 fold: 1- they see the rate of improvement of these models more closely that they felt there is good chance that editors (UI) becomes obsolete in 6-12 months (doesn’t worth the investment) 2- it’s hard to build a UI (editor) that fits everyone (people working with different languages prefer different editors)

What’s wild is how many unexplored AI Agent products are just sitting inside AI models today. These models pack enough intellige…

“You’re absolutely right!” I get this a lot from @claude_code interactions. But I wish I could change it to “You’re kinda right!” Or “That’s a slightly better idea!” Or “That’s an important idea because…” Add variety plz @AnthropicAI

ai agents’ job isn’t to make us productive rather it’s our job to make ai agents productive ai agents are not bound by any capacity constraints in how quickly they can type or the specific hours of the day that they can work or the number of the computers they can access or the number of places they can be at once. so in that world everything flips on its head. therefore enterprise software needs be designed in a way that allows you to: - plan what ai agents should do - to create the tasks that it should be working on - to review the ultimate output that it comes up with We also need to think about how to give ai agents a granular enough task that it will be successful but not so granular that we’re not gaining any real productivity gain. and to ultimately orchestrate and then integrate that work into some bigger picture thing that you are trying to do. great thoughts and mental models from @levie

@thekitze It’s dangerous. Once you try it you can’t go back. The question is do you have enough features/bugs on your desk that you would use it 4+hours a day? If yes then yes.

@thekitze I explicitly stated in https://t.co/UfFe4aeHlk to not db:generate or db:migrate because I may want to remove some fields or make adjustments to Claude’s proposed drizzle schema before migration. But I allowed it run lint:fix and run test after its done with a feature.

AI seems to enable one-person products to take on one-thousand-person products

Building software is becoming 10x faster and 10x cheaper.

Testing pyramid is becoming 10x more important if you want to build scalable software with AI coding agents.

What would you do with a thousand PhDs in your pocket?

Software is changing greatly rapidly. Prompts are becoming the new programming protocol.

“AI is the new electricity.” —Andrew Ng The impactful appliances it will power are yet to be invented.

It’s great time be building in AI and scary time to be investing in AI. Cause an investment may easily fail due to rapid ev…

Fav Quote from report 💭 My fav. quote is in the image below 2 more honorable mentions: “Like electricity and other g…

AI is nitro boost for the impatient only.

We are entering in era of abundance. We are entering in era of chaos. We are entering in era of insane rate of change.…

AI acts as catalyst for productive people to become hyperproductive. It has zero effect on people who have no motivation to…

Many AI apps today feel like the "horseless carriages" of the late 19th century, which "swapped a horse for an engine with…

In AI, we need to be building brand new cars instead of horseless carriages

I don’t know about AGI, but ADI (Artificial Developer Intelligence) seems to be here with the “Claude Code” drop from @AnthropicAI . Just tried it. Not sure if it’s 10X or 100X better than the last best thing. I’m scared and excited. Now we are super developers now?

I don’t know about AGI, but ADI (Artificial Developer Intelligence) seems to be here with the “Claude Code” drop from @AnthropicAI . Just tried it. Not sure if it’s 10X or 100X better than the last best thing. I’m scared and excited. So we are super developers now?

David Sacks Explains How AI Will Go 1,000,000x in Four Years "I would say the rate of progress is exponential right now o…

Github copilot tip for accelerated development of medium to large scale applications: The best way I found to reduce the error rate and unwanted edits and better results while working with copilot: 1. Use "Claude 3.7 Sonnet Thinking" as your model (I think it's best one) 2. Don't select the whole project/folder as the attached context when you want to give an edit task 3. Attach files as context that will serve as examples for the AI on how you want the work to be done (example would be other parts of the app that does handling of something similar) 4. Attach files that most likely need editing 5. Attach files that are relevant for AI to know about to do the task

Github Copilot tip for accelerated development of medium to large scale applications: The best way I've found to reduce the error rate and unwanted edits and better results while working with Github Copilot: 1. Use "Claude 3.7 Sonnet Thinking" as your model (I think it's the best one) 2. Select "Edit" mode 3. Don't select the whole project/folder as the attached context when you want to give an edit task 4. Attach files as context that will serve as examples for the AI on how you want the work to be done (example would be other parts of the app that does handling of something similar) 5. Attach files that most likely need editing 6. Attach files that are relevant for AI to know about to do the task

@salimbinsani Even the best AI coding models such as "Claude 3.7 Sonnet Thinking" still feel like Stackoverlflow on steroids! 💉 These models often write very complex, messy and ugly code 🍝 which then introduces a lot of unhandled edge cases & hard-to-trace bugs 🐞 into existing solutions.

I would rather have AI that does Clean Coding instead of Vibe Coding.

In my case I would go with React because for me: - I know React (: - Way easier to hire React Developer (bigger developer availability in the market) - Big ecosystem and developer community - Typescript is very popular versus Dart - Huge library ecosystem with millions of packages - Uses native components on mobile (native look & feel)

@IcyCortex من مدل غیر فکری اصلا استفاده نمیکنم برای کدزنی

@stats_feed The most impactful ones: - penicillin, - antibiotics, - microscope (we found bacteria&germs), - cleaning water at scale These doubled our life span.

@shl “Software engineer” is too broad. Engineers who gain the most leverage from AI, aka AI engineers, are seeing the biggest…

@chamath As a maxi cursor and replit user, i can share my opinion: - 6500 prompts to generate ~80 modules (python, js) - ~2800…

@chamath If you work at a job as an enterprise software developer that needs to investigate a small bug for 2 days in a large legacy system, you will see how this prediction is not a likely scenario.

AI is turning devs from musicians to DJs.

“The most common error of a smart engineer is to optimize a thing that should not exist.” — @elonmusk Elon Musk’s 5-St…

AI will wildly accelerate those who have good judgement and are highly impatient. Specially if they can have access to capita…

Intelligence is on tap now so agency is even more important

I think if you have ever worked in an enterprise software company (which normally has the biggest portion of the software developer jobs) and had to deal with a large legacy system, your order of difficulty would change. In these sort of companies/systems, investigating a small bug can take days, let alone fixing it.

@IcyCortex یاور هایپ خیلی نویز‌ تولید میکنه. چنتا مطالعه انجام شده در مورد اثر این هوش مصنوعی های جدید مثل چت جی پی تی روی پروداکتیویتی برنامه نویس ها. نتایجش خیلی قابل توجه نبوده.

تجربه من: بین ۲۰-۳۰٪ روز به کد زدن ختم میشه در بهترین حالت. بقیه روز توی جلسه ها، آماده کردن تیکت ها برای اسپرینت بعدی، چک کردن ۲۰ تا کانال توی اسلک و پیام رد و بدل کردن با بقیه دولوپرها، خوندن ایمیلها، چک کردن پول ریکؤست ها، اینوستیگیت کردن باگها، پیگیری کارای جانبی که مدیرم ازم خواسته، تماسهای یهویی برای جواب دادن سوالات مدیر محصول، یا کمک کردن به یه دولوپر دیگه برای حل مشکلش … و خیلی چیزای دیگه. در بهترین حالت این هوش مصنوعی میتونه ۱۰-۲۰٪ توی اون قسمت کد زنی که ۲۰-۳۰٪ کل کارم هست میتونه کمکت کنه. توی کدبیس بزرگ خیلی آسون نیست کمک گرفتن از هوش مصنوعی. چون سخترین کار اینه که کجا باید این کد جدید اضافه بشه و اینکه کجا باید ریفکتور بشه و اینکه چه جاهای دیگه ممکنه نیاز به تغییر داشته باشه… وقتی کد هم بهت میده باید کلی تغییرش بدی خیلی وقتا..،

@shrihacker @StartupArchive_ Same applies here: AI favors the curious mind. If you are not a coder and want to learn, AI can help you greatly. If you are a coder and want to build, AI can help you greatly. If you don’t have the desire. You are left behind whether you know how to code or not.

hate to break it to you too those who are excited, don’t have a good grasp of software development the reality is so far llms maybe can increase up to 10-20% developer’s productivity in “coding” tasks keep in mind developers only “code” around 25-30% of their time but I get it: wishful thinking feels good

It’s mind boggling! I already read Unscripted before. And now I listened to this generated podcast. A few points: - This idea of turning books to short deep dive podcasts can already make businesses like “Blinkist” go down 😯 - I wish I could talk like these AI hosts 😂

“AaaS” 😅 or Architecture as a Service… this is what I called a project that a friend of mine did in a weekend and showed it to me tonight. It was pretty impressive, I will be his first user. Will clone the code, choose my database and type of application and boom…I will have authentication, authorization, extremely fast linting and testing library, api validation, docker file that generates under 200MB image, ci/cd pipeline, automated scripts to deploy to cloud services, orm, rpc, logging, tracking, monitoring, schema migration automation… 🤯 This can be used to start a scalable full stack application in an extremely fast way…

@IcyCortex براش نقل قول مارتین فولر رو بفرست 😂 “Any fool can write code that a computer can understand. Good programmers write code that humans can understand.”

Big tech taxes innovation and creativity! and it's not a low tax 🫤 I really don't like the economic state of the internet today... Most of its economics is flowing through gateways of a handful of big tech companies and they are taking a ridiculous percentage of it. Apple & Google app stores: ~30% YouTube: 45% Instagram, TikTok, Facebook... ~100% 🤯 The big tech is making hundreds of billions off of users attention and engagement, creator's creativity, and innovator's ideas and inventions! Again: they tax innovation and creativity! 😡 When there is low economic incentives to be creative (because of this big tax) and … when you know you have no power over these giants and they can wipe out your creations or tax you even more, innovation slows down 😢 This book made me at least 10 times smarter about internet, web, protocols, networks, economics and innovation. It's the education of a lifetime if you care about any of the above or human progress in general. @cdixon brilliant analogies and perspective! I ordered 3 more copies to give as gifts! @cdixon @cdixon explains how blockchain networks are essential building blocks of the next era of the internet and how they can counterbalance the current internet consolidation by big tech Happy reading! 📙 #bookreview #web3 #innovation #blockchain #crypto #creativity #reading @cdixon Find my notes in the booksense app. Link in the bio.

Technology is what makes us human. It's what distinguishes man from animal. Fire arguably made us human. The invention of fire allowed humans to outsource some of our metabolism to the fire and allocate more scarce calories for brain development. Truth and technology go hand-in-hand. Technology is birthed everytime we discover a non-obvious truth. A truth either about humans or the world. Build products based on truths other people can't grasp. If it works, they will buy it. Their confusion is your advantage. Unlock unseen value. It requires more empathy and more imagination to think about value that isn't seen. You can see a skyscraper. You can't see what could've been built but wasn't. If the internet was programmable communication, crypto is programmable money. Before the internet, you needed a deal with a telecom company to deploy information-transferring code. Before Bitcoin, you needed a deal with a bank to deploy value-transferring code. Those saying, "Crypto is just another asset class" sound like those who said, "The internet is just another media outlet." They didn't understand programmability, permissionless-ness, or peer-to-peer, and they overestimated the robustness of legacy institutions. @balajis is a scientist, technologist, entrepreneur, investor, and the author of a recently published book "The Network State". If the internet was programmable communication, crypto is programmable money. Before the internet, you needed a deal with a telecom company to deploy information-transferring code. Before Bitcoin, you needed a deal with a bank to deploy value-transferring code. Those saying, "Crypto is just another asset class" sound like those who said, "The internet is just another media outlet." They didn't understand programmability, permissionless-ness, or peer-to-peer, and they overestimated the robustness of legacy institutions. @balajis is a scientist, technologist, entrepreneur, and investor who was previously CTO of Coinbase and the author of a recently published book "The Network State: How to Start a New Country". @balajis @EricJorgenson writes and podcasts about technology, startups, and investing who previously published one of my favorite books: "The Almanack of Naval Ravikant". Happy reading! 📙 @balajis @EricJorgenson Find my notes in the booksense app. Link in bio.

How web3 data portability reduces the power of centralized services 🧵

@NavalismHQ @naval There is a shortage of developers. Actually its a big one.