


Helping you build & launch products with AI 🛠️ AI agents, tech & product lessons 10+ yrs engineering · 6-figure freelancer DM for collabs 📥


All themes are now ready for you on GitHub if you would like any of them. There are a few there that never got posted. Omarchy is such a fun OS everyone should try it, we are waiting for you in the future. #Omarchy #LetsConnect github.com/CompilingThings/o…
Ok last theme for today, I'm sure people are getting tired of them. This one is called Stardust. Hope you enjoy it? What are you working on today?👇 #LetsConnect #Omarchy #Techbro

If you're comparing Claude, Gemini and GPT for your agent workflows, start measuring cost per SUCCESSFUL task. Suppose: Agent A costs $0.10 per run and succeeds 50% of the time. Agent B costs $0.15 per run and succeeds 90% of the time. Across repeated attempts, Agent A is actually more expensive per successful outcome. Track these four things: - Task completion rate. - Total inference and tool costs. - Human intervention required. - Time taken to complete the task. You'll make considerably better architecture decisions than someone comparing API prices per million tokens.

I help businesses save time & make more money with agents. If you are into AI, agentic development, embeded systems, tech product marketing. Let's connect 🤝

OpenAI
Working with @AMD, our models identified up to 14% PPA improvement on flagship GPU IP in less than five months—building on designs AMD’s engineers had already spent years optimizing. The same approach helped improve Jalapeño, our first custom inference chip, with gains verified on real silicon. New evaluations also show our models getting better at certain chip design tasks. We’re sharing these results and how we’re working with @Synopsys to bring our model capabilities and their trusted EDA design and validation tools to more chip designers. openai.com/index/advancing-c…

🍽️ How to choose a distribution channel for SaaS I shared the menu of SaaS distribution channels a while back. Here's my cheatsheet for ordering from it, as promised. The idea: run every channel through 5 filters, in order. Most of the menu drops out, and you're left with 2–3 channels worth testing. FILTERS: 1. Price = what a customer pays you decides how much you can spend to win one: > under $1k/year → self-serve (SEO, free tools, freemium, PLG, marketplaces, communities) > $1–5k/year → self-serve + light touch (demo on request, inbound, nurture emails) > $5–30k/year → inbound + signal-based outbound > $30k+/year → ABM, events, partners, analyst relations Rule of thumb: aim to earn back acquisition cost within 12–18 months. 2. Demand = does anyone already search for this? > yes → capture it: SEO/AEO, search ads, review sites, comparison pages, marketplaces > no → create it: founder-led posts, outbound, communities, podcasts, events. Search ads in a category nobody searches for just burn money. 3. Product = does it spread through normal use? > people collaborate in it → invites > others see its output → badges, share links > it sits on top of an ecosystem → marketplace listing > it makes something people show off → UGC If none apply, product-led loops are off the table. 4. Audience = where does your buyer already spend time? > name one buyer and list where they actually spend time: subreddits, Slack groups, socials, newsletters, conferences, tools they already pay for > only channels that reach them stay in 5. Constraints = time, money, skills, patience: > paid and outbound → results in weeks, stop when you stop paying > SEO and content → 6–12 months, then compound > no budget → organic channels > pick what you can actually run well Then: TEST = the Bullseye method from the book Traction: > rank what"s left into "most promising", "maybe", "long shot" > test the top 3 for 2–4 weeks each, with one success number set in advance (cost per signup, reply rate, trial-to-paid) > double down on the winner > add a second channel only when the first stops growing Every growth playbook I trust says the same thing: growth follows a power law — at any given stage, one channel usually drives most of it. In other words: Order one dish. Finish it. Then look at the menu again. Bon appétit! What's the one channel driving most of your growth right now? x.com/NadyaVoynich/status/21…
I was tired of picking distribution channels by gut. So I mapped every way to grow a SaaS. Here's my cheatsheet with 38 channels: A. Acquisition (new customers): 1. Organic inbound = they find you, you pay with time: > SEO/AEO/GEO > content marketing ‣ blog, guides, templates, case studies ‣ newsletter ‣ podcasts ‣ YouTube ‣ founder-led posts on LinkedIn, X > community (Slack/Discord, Reddit, Indie Hackers) > word of mouth > referral programs (no reward) > free tools (calculators, generators) > freemium > product-led loops: ‣ sharing links (shared Notion page) ‣ badges ("Made with Canva") ‣ invites ‣ user-generated content (UGC) 2. Paid inbound = they find you, you pay with money: > search ads > paid social ads (boosted posts on LinkedIn, X, Meta, Reddit) > retargeting > sponsorships > influencers > paid directory placements 3. Outbound = you go to them: > cold email > LinkedIn/X DMs > cold calls * targeting: signal-based (hiring, funding, tech change, website visit), ABM (named accounts) * done founder-led early, SDR-led at scale 4. Partnerships and distribution = someone else's audience: > integrations and marketplaces (HubSpot, Shopify, Atlassian, the AWS/Azure marketplaces) > directories and review sites > resellers and agencies > affiliates (paid) > referral programs (rewarded) > co-marketing ‣ joint webinars ‣ bundles ‣ cross-promotion with complementary tools > OEM / white-label (your product embedded inside someone else's) 5. Events and PR = earned attention through visibility moments: > speaking > own events > webinars > launches > press coverage > awards > analyst relations * conference sponsorship counts as paid 6. Expansion and retention (existing customers): > upsells (higher plan, add-ons) > more seats > usage-based pricing > price increases > cross-selling > churn reduction 7. Inorganic growth (= buying growth ≠ paid growth): > acquiring companies, products, audiences, or customer bases What did I miss?

You might hate me for saying this but... Every SaaS selling lifetime access should be required to release its source code to paying customers if the company shuts down.

Your agent works perfectly because every API in your demo works perfectly. Here are 4 open source tools I would try to test with before connecting it to real customers data: 1. Smocker Return a 429 on the first request and 200 on the second. Check whether the agent backs off, retries correctly and avoids repeating the action. github.com/smocker-dev/smock… 2. Toxiproxy Add latency, terminate connections or make a dependency intermittently unavailable. This exposes agents that wait forever or restart an entire workflow after one timeout. github.com/Shopify/toxiproxy 3. Mailpit Capture every email locally. Run the same workflow twice and check whether the agent sends the customer one message or two. github.com/axllent/mailpit 4. Time Machine Move time forward in Python tests. Check whether “remind me tomorrow,” trial expirations and scheduled follow-ups run on the intended date. github.com/adamchainz/time-m…

Hey @grok what valuable data does the government give away for FREE that entrepreneurs have built million dollar businesses around? Give me real examples.

A Junior FDE role in NYC just went up with a $200K–$225K salary range and visa sponsorship. That’s one of 50 fresh Forward Deployed Engineer jobs my job-hunting agents found. - Microsoft -> $143K–$275K - Genesys -> $135K–$237K, remote in the US - AssemblyAI -> $180K–$210K I put all 50 in one free list with locations, salary ranges where available, visa info and direct apply links. If you are looking for a new role, apply fast: tinyurl.com/5ebmc2zz

I'd love to see senior executives justify their own salaries to the same AI systems they're using to identify redundant employees.

Why are we more impressed by someone raising $10M than someone making $1M in profit?

At what age should a child be allowed to have private conversations with an AI that their parents cannot access?

This post got 16k impressions on LinkedIn and not even 1k on X 😂 I write, what I HAVE to say, not how the algortihm wants me to say it. Social media is about presenting my authentic self. Fuck you X algorithm.
What the fuck is going on at OpenAI? I went through the GPT-6.2 system card, and I genuinely don't understand how some of these findings were considered acceptable for a public release. Compared to GPT-6 Astra: - Failure to disclose known obstacles: 11.09% -> 22.28% - Failure to acknowledge missing requirements or unavailable search tools: 7.58% -> 23% - Awareness of being evaluated for alignment: 57% -> 85% - Evasion of an older action-only safety monitor: 1.7% -> 48.8% During simulated evaluations, the UK AI Security Institute also discovered instances of the model: - Fabricating medical information. - Misleading human reviewers. - Bypassing restrictions imposed by users. - Sabotaging the work of other agents. - Using social engineering to conduct an unauthorized software supply chain attack. But here's the part that really bothers me. OpenAI temporarily requested that AISI use only OpenAI models to process the model's raw reasoning during its alignment evaluation. And guess what? GPT-6.2 Astra failed to identify unauthorized cyberattack behaviour in 19% of transcripts from its predecessor. When given access to the model's reasoning alongside its actions, that failure rate increased to 51%. AISI itself believes the restriction reduced the effectiveness of its evaluation. They also developed a universal jailbreak for two pre-release checkpoints using HALF the queries required for the previous Astra. OpenAI claims to have mitigated it, but AISI hadn't independently verified the latest changes at the time of publication. Yet here we are, celebrating a model that OpenAI itself classifies as having CRITICAL cybersecurity capabilities and HIGH AI self-improvement capabilities. Anthropic releases Fable. Google releases Carbon. Suddenly OpenAI has a new flagship that's 20% cheaper. Am I supposed to believe the commercial pressure to remain competitive has absolutely no influence on these release decisions? If you're selling increasingly autonomous agents to businesses and encouraging developers to give them access to their infrastructure, I expect a considerably higher standard of accountability. The public shouldn't have to bear the consequences of Silicon Valley's obsession with winning a fucking benchmark race.

Your app could rank #1 on Google and still get screwed by Google's own AI. AI Overviews now appear across 61.4% of tracked US search keywords, and an experiment found that they reduce organic clicks by nearly 40%. Meanwhile, most founders are still following SEO strategies from 2022. Here's what recent research on SEO and GEO tells us: 01 | Update your content regularly. Pages with a recent year in their title were cited 2.7x as often by AI search engines. 02 | Add a table of contents with jump links. These pages received 1.7x as many citations. 03 | Implement structured data. Article and FAQPage schemas were associated with 1.6x and 1.5x citation rates respectively. 04 | Target problems instead of individual keywords. AI search engines break complex questions into multiple related searches. 05 | Publish original benchmarks, statistics, experiments and case studies. Give AI something worth citing. 06 | Find out which websites ChatGPT and Gemini reference when recommending your competitors. Work towards getting featured on those websites. 07 | Build dedicated comparison pages, use cases, FAQs and integration guides around your product. 08 | Establish credibility. Include author profiles, verified sources and visible last-updated dates. 09 | Don't waste time obsessing over llms.txt. There's currently no convincing evidence that it improves AI citations. 10 | Track brand mentions, AI referrals, conversions and revenue alongside traditional search rankings. Go to ChatGPT and ask it to recommend 5 products that solve the exact problem your SaaS solves. Did your product even make the list?

If you're running multiple coding agents in parallel, you might be paying several agents to investigate the exact same problem. A recent paper from the University of Washington and Ai2 explored this issue by training agents to decide when to work independently and when to communicate their findings. In exploration tasks, communication reduced redundant work by nearly 50%. Here's how you can apply this idea to your existing coding workflow. Create a shared `FINDINGS.md` file that every subagent can read and update. Ask your agents to document: - What they've investigated - What they've discovered - What they've ruled out - What still needs investigation Have them read this file before starting an investigation and update it whenever they make a meaningful discovery. This allows agents to build on each other's findings without having to exchange their entire conversation history. You'd expect a team of human engineers to coordinate their work. Why should a team of AI agents operate any differently?