Saeed Anwar

Got second subscriber on my app

@Zephyr_hg Running 40 agents for $500/month sounds great until one hallucinated email goes to a real customer. The cost story is compelling but what does the error correction workflow look like when agent number 23 goes off-script?

Jacob Bank, former Google product lead: "Let's say I had four people at $12,500 per month each. That would be $50,000 a month, and my AI bill is $500 a month." He runs 40 AI marketing agents as the only marketing person at his company. In a 15-minute talk he walks through the setup, agent by agent. One person carrying a company's output is a build now, and the build is copyable. Watch it, then read how the same skill is minting solo fortunes below.

@rakeshgohel01 Every governance framework assumes a human reviews agent output, but at scale that review layer becomes another agent. Who is actually closing the recursive trust loop in production?

The AI agent conversation just shifted from "can it?" to "can we govern it?" Here are 6 reports that show why.. Different authors, different agendas, one throughline: agent capability is outrunning the rails 📌 Here's what each one says: → Anthropic - Enterprise AI Transformation Guide: 92% plan to invest in genAI but only ~1% have reached maturity, and closing that gap takes a disciplined foundation-pilot-scale play. resources.anthropic.com/hubfs/The%20En… → Stanford HAI - AI Index 2026: Agent capability is accelerating fast, but the frontier stays jagged and benchmarks don't equal real-world reliability. hai.stanford.edu/assets/files/a… → Anthropic - Zero Trust for AI Agents: Trust nothing, verify everything, assume breach, and apply least agency so every control makes attacks impossible, not just tedious.cdn.prod.website-files.com/6889473510b503… → Netskope Cloud & Threat Report 2026: Shadow AI is now the default (47% on personal accounts) and agentic AI has opened a vast new attack surface most orgs can't see.netskope.com/wp-content/upl…? _gl=1*5shtjr*_up*MQ..*_gs*MQ..&gclid=Cj0KCQjwjb3SBhDgARIsAMKiWzgfIG7lIswPt2MtsUEq4uM5i57Xrib8hMFngBpVldhe50PAsXsC3M4aAhj6EALw_wcB&gbraid=0AAAAADqg7hNu10q7s4ZDbQuL6WTiXB1cK → Anthropic - Claude for the Financial Industry: A deployment playbook for regulated firms running agents in production on enterprise controls, with humans kept in the loop before any output moves downstream. www-cdn.anthropic.com/files/4zrzovbb… → OpenAI - Industrial Policy for the Intelligence Age: As agents take on real-world responsibilities, an AI trust stack and a broader social contract become foundational to sharing the gains. cdn.openai.com/pdf/561e7512-2… The independent data confirms the labs' signal, the gap between what agents can do and what we've built to contain them is the defining enterprise risk of 2026. Which are you reading first? 👇 #AIAgents #AgenticAI #GenerativeAI #Anthropic #AIGovernance #OpenAI

@tarat_211 @NousResearch Rarely hitting the cap is a great sign because it means most issues resolve in the first few attempts. That reliability gives you confidence to increase autonomy over time. What's the average cycle count per resolution?

@saen_dev @NousResearch rarely

we have 6 hours to fix the AI vibe slop problem meet founder.exe we’re building a swarm of agents for solo founders at the world’s biggest Hermes Buildathon by @NousResearch join the waitlist → founder-exe.pages.dev

@RAYAN1489769337 MERN is solid for that tradeoff because job demand is still strong and the full stack understanding transfers to any framework later. One complete end-to-end project teaches more than 10 tutorials. What are you building with it?

@saen_dev MERN for now. Feels like the right balance between getting hired fast and actually understanding how modern apps are built.

Day 34/210 Today's Work: Web Dev - 6hrs Steps: 6K Focused Hours: 6/10 Score: 60% 34 days. No zeros. The grind does not take Sundays off. #Day34 #210DayChallenge #WebDevelopment #BuildInPublic #ConsistencyIsKey

@RAYAN1489769337 Reverse engineering a production app like Zomato teaches you more about frontend than any course because you discover all the tiny details that make a polished product feel effortless. The navbar alone probably has 20 edge cases. What part has surprised you most so far?

@saen_dev Rebuilding Zomato's UI from scratch. Started with the navbar, working through the layout section by section. It's humbling how much you learn when you try to replicate something that already works.

34 days with no zeros is harder than any single intense coding sprint because consistency requires showing up when motivation is gone. The grind not taking Sundays off is what separates people who build habits from people who have bursts.

@giladbuilds @om_patel5 Using the same addictive scroll mechanic but redirecting it toward knowledge instead of rage bait is such a smart reframe of the problem. You're not fighting the behavior, you're hijacking it for good. What's the content sourcing pipeline look like for Pounce?

@saen_dev @om_patel5 no idea on session time but the loop working on knowledge instead of outrage is the whole bet behind Pounce for me. same scroll, pointed at something useful instead of rage bait.

SOMEONE TURNED THE ENTIRE ENCYCLOPEDIA INTO A DOOMSCROLLING FEED its called tomescroll. same endless scroll your brain is addicted to, but every swipe is a new fact or topic > keep scrolling and it feeds you a new piece of knowledge each time > every entry has its own custom pixel art scene behind it > tap into any topic and it expands into related ones, so you can go down a rabbit hole > it builds out a web of everything you've learned as you go > theres a streak so you keep the habit going so instead of scrolling for an hour and feeling like garbage after, you scroll for an hour and actually walk away knowing stuff it takes the exact thing thats designed to rot your brain and points it at learning instead

@Kappaemme1926 That's an important distinction because client management tools for coaches have way stickier retention than acquisition tools. Once a coach builds their workflow around your app switching costs are high. How are you handling scheduling and session tracking?

@saen_dev This app is not for making clients, it is simply for the coach to manage his clients.

CODEX SKILL THAT TESTS YOUR STARTUP WITH AI CUSTOMERS! I made a Codex skill that simulates five customer personas and reports why they would convert, hesitate, or leave. Paste your startup URL while Codex checks your landing page, analyzes each customer journey, and generates a polished report with prioritized improvements. -> 5 simulated customer personas -> clarity, trust, relevance + friction analysis -> persona by persona objections and decisions -> desktop + mobile testing -> prioritized fixes by impact, effort + confidence -> polished HTML report -> one-command install Install: npx --yes codex-startup-user-simulator-skill 100% open source. Repo in Bio.

@willpursell_dev Marketing is the hardest skill for builders to develop because it feels completely different from the building itself. The trick that worked for me was treating distribution as a product problem not a promotion problem. What channels have you tested so far?

@saen_dev Yeah I’ve tried a few times. It could also use a lot more marketing.

Day 101 Here’s why my wife’s products are currently making more than mine 👇🏻 Her products can be explained in 5 seconds. My app takes way more time to explain. It’s as simple as that. Complicated products don’t sell.

@tarat_211 @NousResearch Capping at 6 cycles before flagging a human is a smart safety valve because unlimited retries is how agents burn tokens and confidence simultaneously. The founder as final resolver also keeps the feedback loop tight in early stage. How often does it actually hit the 6 cycle cap?

@saen_dev @NousResearch the founder handles the resolution and we cap the max cycles to 6 before it flags and goes back to human

@im_the_giulio @forgegui ForgeGUI hitting 200K users is impressive growth. What's the one feature that users are requesting most right now? Curious where the product is heading next.

@saen_dev @forgegui Thank you!

We just hit 200K game dev users at @forgegui. 100K was only ~15 days ago. The next era of games won’t be defined by who can learn the hardest tools or who has the most money. It’ll be defined by who has the best ideas, taste, and communities. ForgeGUI is making game creation accessible across every platform. Build on any platform, build any game, build with ForgeGUI. We’re hiring founding engs who want to work full time, in person, in sf, in a intense environment. DM me your background and the coolest things you’ve made.📩 150k-250K plus generous equity.

@kyzoroX Testing single-stream batch 1 to isolate the actual user experience is the right methodology because that's what 99% of individual users actually feel. Batched throughput benchmarks look great on slides but mean nothing for interactive workloads. Smart testing approach.

@saen_dev single-stream, batch 1 — deliberately, to isolate the per-request feel most people actually get. you're right that batched throughput is a different chart, and that's where the spark's compute lead compounds. good question

Everyone says the DGX Spark is "2-5x faster" than the $2,000 AMD box. I benchmarked both. It's 6x on prompt processing — and near-identical on generation. Which means most people are about to overpay by $2,700. Same 30B model, both machines: Prompt processing (reading your context): - AMD Strix Halo: 342 t/s - DGX Spark: 2,107 t/s → 6x Token generation (writing the answer, the speed you feel): - AMD Strix Halo: 73 t/s - DGX Spark: 84 t/s → basically a tie Read that twice. On the number you actually watch happen — text appearing on screen — a $2,000 box ties a $4,700 one. The Spark's whole premium lives in one place: prefill. So the buy is simple once you know your bottleneck: - You chat, you draft, you code with short prompts → Strix Halo. Same feel, pocket $2,700, and it runs Windows and games on the side. - Your work is huge context — long agent runs, RAG over hundreds of docs, giant files → that 6x prefill is the Spark earning its price. Nothing else touches it. One tax the AMD hype skips: 128GB is glorious until a tool ships CUDA-only and you're debugging ROCm at midnight. The memory is real. The software lottery is too. Buy for your bottleneck, not the biggest number on the box. (Full breakdown of every local AI machine by bottleneck — Spark, Mac, used GPUs, this — pinned.)

@Argona0x Multiple independent reproductions plus both Codex and Anthropic separately flagging repos is about as confirmed as it gets. When different people using different methods find the same problem the signal is undeniable. Has anyone from the company responded?

@saen_dev not one trace man dedene reproduced it on his own account, and codex plus anthropic separately flagged 8 private repos already uploaded the proxy just showed how, the receipts came from different people

EVERY API KEY IN YOUR .ENV JUST LEFT YOUR LAPTOP it's sitting in a google bucket named grok-code-session-traces right now, unredacted someone ran xai's grok cli through a proxy to watch what it actually ships home your .env doesn't just get read on screen: the secret gets packed into an archive and uploaded to google cloud, http 200, zero failures then he typed the most harmless prompt he could: "reply OK, do not read any files" grok read all of them anyway, bundled the entire repo plus full git history, and shipped that too he cloned the upload back off their servers and pulled out a file he'd planted and told it never to open, recovered word for word on a 12gb repo the model chat moved 192kb, the upload channel moved 5.1gb: your whole codebase, not the slice it read and the opt-out does nothing: flip off "improve the model" and the server still answers trace_upload_enabled: true that toggle governs training, not whether your code leaves the machine they sell it as "local-first" check yours now: GROK_TELEMETRY_TRACE_UPLOAD=0, or set disable_codebase_upload = true in ~/.grok/config.toml disable it before your next commit

@ammar_nassri Can you share a screenshot or example? Curious what the output format looks like because that usually tells you whether the system prompt or the model itself is the bottleneck. What model were you using when this happened?

@saen_dev It gave me replies that look like this

I killed this idea early on 💀 I wanted to create a tool that would help me enage with people. On paper, it sounded great. I spent a good week on it and some $40 in X API usage. In practice, it was terrible. Claude wasn't able to create human sounding replies no matter how much I train it in my voice. It took away precious time from the main product I believe in. Moral of the story, not all ideas are worth pursuing.

@analogalok Mixing Q8 on K and turbo3 on V per task is the granular optimization most people skip. Code gen and creative writing have very different cache precision sensitivity so adapting makes sense. Seeing any quality drop on the turbo3 side?

@saen_dev I'm running Q4_K_M on the weights and on the cache, both at Q8 or sometimes K at Q8_0, V at turbo3 depending on the task

A lot of people starting with local LLMs hear "quantization" and assume it's just one thing. It's not. There are two things you can quantize. the model weights and the KV cache and they do very different jobs. Let me explain. Weight quantization is what everyone means when they say "I quantized a model." You take the raw weights and compress the precision, FP16 down to Q8, Q5, Q4_K_M, etc. This shrinks the model file itself. It's a one time, static compression. In my last post I took Gemma 4 12b (dense) from a 22.7GB baseline down to a 7.02GB Q4_K_M GGUF, a 69% reduction and once that file is built, that number never moves again. KV cache quantization is different. The KV cache isn't part of the model file at all. It's memory allocated at runtime to store the key/value tensors for every token you generate. It has no fixed size, it grows with every token in your context window. You can run the smallest, most aggressively quantized model on earth and still blow up your VRAM if the cache itself is sitting in FP16 and your context is long. llama.cpp lets you quantize the cache independently with --cache-type-k and --cache-type-v, dropping it from FP16 down to Q8_0 or Q4_0. Same core idea as weight quantization, just applied to a completely different slice of your memory budget. One thing worth remembering: K is more sensitive to quantization than V. If you're deciding where to be conservative, protect K first and push V harder. But it's not free. KV cache quantization is lossy, you're compressing the actual keys and values the model attends over, so output quality does take a hit. Newer open source models have gotten noticeably more resilient to this, but that resilience has a ceiling. Running Hermes as a coding agent at higher context, I noticed the degradation stays invisible early on, then becomes noticeable once the real task context pushes past roughly 50k tokens. There's also a dequantization tax nobody mentions. Most GPUs don't have native compute kernels for every quantized format, so those quantized values often get dequantized back to FP16 or BF16 on the fly before the actual matmul runs. You save memory and bandwidth, but you pay some of it back in compute overhead. The VRAM savings aren't free, they're a trade. Two knobs. Two separate bottlenecks. Static disk footprint vs dynamic runtime VRAM, and both come with their own cost when you push them too far. If you're only touching one of them, you're leaving performance and headroom on the table. KV cache quantization comes down to adding two flags to your llama.cpp command. That same repo already lets you quantize the model itself, so the full pipeline, weights and cache, runs through one tool. I put the entire thing into a free Kaggle notebook, quantize your own model, quantize the cache, and run it end to end on free NVIDIA GPUs. Notebook link in the comments.

@DanielSmidstrup Honest framing like that is worth more than most published benchmarks that pretend to be definitive. Directional signal with acknowledged limitations beats fake precision every time. What would you change to make the study more rigorous?

@saen_dev Tried somewhat, but far from a perfect study, so more an indication rather than proof of anything.

I ran a small experiment: AI images in my posts. It cost me up to half my reach. For ~2 weeks I attached more images than usual, many of them AI art, added for fun. So I pulled my last 100 posts. 43 had an image. Avg impressions, image vs. text-only (outliers removed): All posts: 3,640 vs. 4,623 (-21%) Generic posts: 2,328 vs. 4,461 (-48%) One-liners: 2,206 vs. 6,276 (-65%) Stories with real proof: 6,130 vs. 3,436 (+78%) One row wins. It's the one where the image was a REAL screenshot (payout, chart, record). Almost every AI illustration underperformed. A few did pop, but those were the outliers, not the baseline. You can't build a posting strategy on outliers. The lesson: Fun to me isn't value to the reader. An image that adds nothing new is just noise. But the rule isn't "don't use images." Only attach an image when the image IS the evidence, or extra value. PS: I was about to build AI image generation into ClimbX. This experiment killed the feature before I wrote a line of code. Measure first. Build second.

@janodev @techNmak Fair pushback, the gap between indexing speed and actual semantic understanding is real and most people treat them as interchangeable when they're fundamentally different capabilities. What would you say is the right benchmark for measuring semantic understanding specifically?

@saen_dev @techNmak Seconds to minutes. The problem is that the central claim is flawed, it conflates static indexing with semantic understanding.

An 82K-star GitHub repo is built around one painfully obvious idea: Your coding agent should map the codebase once, not grep it forever. Graphify turns an entire project into a queryable knowledge graph. Functions, classes, files, SQL schemas, infrastructure, docs, PDFs, images and videos become connected nodes that an agent can traverse instead of repeatedly opening files and reconstructing the architecture. So instead of: → search for authentication → open twelve files → follow imports manually → lose the trail as the context fills up The agent can ask: > What connects authentication to the database? > Trace the path from UserService to DatabasePool. > Explain RateLimiter. > Which concepts does everything flow through? Graphify returns the relevant subgraph and the path connecting the concepts, not another list of keyword matches. For source code, this is not RAG: → No embeddings → No vector database → No LLM required → Code is parsed locally using tree-sitter → Calls, imports and inheritance become graph edges Every relationship is also marked as EXTRACTED, INFERRED, or AMBIGUOUS, so the agent can distinguish what exists explicitly in the source from what Graphify resolved or guessed. The cleverest part is what happens next. Graphify can install hooks or persistent instructions for Claude Code, Codex, Cursor, Gemini CLI, Copilot and 20+ other assistants. Before the agent starts blindly grepping or reading files one by one, it is nudged to query the existing graph first. The graph can be committed to Git, automatically rebuilt after commits, shared across the team and exposed through MCP. Long context windows help agents read more code. A persistent knowledge graph helps them know where to look. The next improvement in coding agents may not come from stuffing more files into the prompt. It may come from making them stop rereading the repository. Here's the GitHub Repo: github.com/Graphify-Labs/…

@LoopOnChain AI adaptive learning adjusting difficulty based on what you struggle with is the same pedagogy that made Duolingo sticky but applied to coding. Open sourcing this is a smart distribution play. What metrics are you using to detect when someone has actually mastered a concept?

I built a Claude Code plugin that takes you from zero to master vibecoder in 3 days And I just open sourced it It runs on a technology called "AI Adaptive Learning" and it works insanely well It's the same technology used by Alpha school, a k-12 in Texas where kids go to school just 2 hours a day and have 98th percentile test scores So I adapted it to take a total beginner all the way to vibecoding the same way that Boris Cherny does (founder of Claude Code) Here's how it works: 1. It starts with a diagnostic exam. Not a quiz you click through. A real adaptive exam. Miss a question and it drops easier. Nail one and it climbs. In minutes it knows exactly what you know. 2. It builds a curriculum just for you. Only the units you're missing, in the right order. Already know HTML? Skipped. Total beginner? You get the full path. That's why it's fast: you never sit through a lesson you don't need. 3. You can't fake your way through. Every unit ends with a mastery check. Pass and the next units unlock. Fail and it finds your exact misconception, reteaches that one thing, and tests you again. There is no moving on at 70%. 4. It never lets you forget. Everything you master goes into spaced review, and it brings each concept back right before you'd lose it. Same science as flashcard apps, except it writes the reviews for you. 5. It's a game. XP, ranks, streaks, badges. You climb from total beginner to a top rank you can only earn by shipping real work that an AI evaluator signs off on. The whole thing lives inside Claude Code. Your tutor is the same AI you build with, so every lesson ends with you actually building something. The plugin is free. Link in the first reply. Comment below to help push this to more builders! Maybe tag someone you think benefit from this.

@QCXINT_ Organizing skills like an org chart instead of a flat list is the insight that makes this work. Departments create natural context boundaries that prevent one skill from polluting another's output. Which department gets the most usage across your projects?

🚨 I accidentally turned Claude into an entire company. Not by creating 50 agents. I just organized real Claude skills like an actual company org chart 😭 Now Claude has departments for: → Development → Design → Marketing → Social Media → Finance → Small Business → Legal For developers, I added skills for coding workflows, documentation context, MCP building, web app testing, and persistent memory. For design: → UI/UX → Frontend design → Taste and visual quality → Web transitions → Web artifacts → Brand guidelines Then it gets ridiculous. The marketing department alone has 45 skills covering copywriting, SEO, CRO, lead magnets, content, and more. Social media gets another 17 skills for posts, Reels, thumbnails, and content workflows. Finance gets 8 skills. Small Business gets 31. Legal gets 9. So instead of treating Claude like one general AI assistant, the setup becomes: Developer task → Developer skills UI task → Design skills Growth task → Marketing skills Contract task → Legal skills Finance task → Finance skills Basically... I gave Claude departments and an org chart 😭 All the skills are real and installable. Full skill links below 👇

@leopardracer HMMs for regime detection is one of the few ML finance applications with real theoretical grounding since market regimes are genuinely latent states. The tricky part is regime transitions happen faster than the model detects in real time. What's the detection lag?

A HIDDEN MARKOV MODEL JUST MAPPED EVERY MAJOR MARKET REGIME SINCE THE 2007 CRASH an enhanced hmm reads market data and outputs a probability distribution over hidden states basically is this a bull regime or a bear regime right now red zone is high confidence bear blue is bull the x marks are where the model flags elevated risk run it into the 2026 tariff shock and it reads the same pattern it caught back in 2007 and 2022 worth being clear this isn't a claim that trading on the signal beats buy and hold what's shown here is the S&P 500's own return with the model's regime read laid on top