Lumina Strategy

583 posts

Lumina Strategy

@LuminaStrategy

We build AI systems that drive revenue and reduce cost. We only win when your business wins. Practical AI. Real results.

England, United Kingdom Katılım Haziran 2024

99 Takip Edilen141 Takipçiler

Sabitlenmiş Tweet

Lumina Strategy@LuminaStrategy·4 Nis

Two heavyweights publicly changed their minds on the same day. @BrianRoemmele just admitted he was wrong about Google’s “no moat” stance back in 2023 and praised Gemma 4 running locally as the real play. @garrytan just flipped on MCP: “I changed my mind. It can be wonderful” if you stop treating it like a shitty REST shim and build it right for agents. This is the signal. Not that they flipped. But that the top operators in the game are updating in public, in real time, as the ground shifts under everyone. The people who win from here aren’t the ones who were right in 2023 or 2025. They’re the ones willing to be wrong today and ship anyway. Intellectual honesty at this speed is the new unfair advantage.

English

5.9K

Lumina Strategy@LuminaStrategy·23 Nis

@BrianRoemelle This is brilliant. Turning big accounts into active talent scouts with real selfish incentives (multipliers, badges, revenue share) is exactly what X needs. Shifts the game from algo lottery to merit-based discovery. Game-changer. This deserves to be built. @elonmusk

English

Minocrisy@MinocrisyHQ·20 Nis

Instead of complaining to @nikitabier it @elonmusk about the algorithm I decided to fix it. I was chatting to @grok and this is what we came up with. The whole platform is built around the feed as the main discovery engine, so posting on your own profile feels like screaming into the void unless the algo decides to bless you. Huge accounts (the ones with real reach) almost never manually scroll into random profiles anymore. They’re drowning in notifications, replies, and their own For You tab. Small creators are basically farming impressions that never come because no one with power ever visits the page. A winning incentive system would flip that behavior by making it profitable or status-rewarding for big accounts to actively hunt smaller ones. Here's a few versions that could actually work: “Scout Multiplier” program: Verified / high-follower accounts (say >100k) get a temporary impression multiplier on their own posts if they spend real time on smaller profiles (<50k followers). X could track it via time-on-profile + likes/replies/bookmarks. Hit the threshold for the day? Your next 5 posts get 2–3x normal reach. It turns talent scouting into a selfish growth hack for the big accounts. Discovery Quests + Leaderboard Daily or weekly challenges only visible to large accounts: “Visit & engage with 10 accounts under 20k followers → unlock a blue-check ‘Talent Scout’ badge + bonus ad revenue share for the week.” Public leaderboard of “Top Scouts” (with follower count and number of small accounts boosted). Ego + money = powerful combo. The badge itself becomes social proof. Reply-to-Boost Chain: If a big account replies to a small one, the small account’s entire recent thread gets a one-time algo super-boost (visible to the big account’s audience). The big account gets credited publicly ('@BigAccount just amplified @SmallAccount — now 187k saw it'). Creates a visible 'I found this gem' flex culture. Revenue share for the scout: take the existing creator ad revenue program and add a “referral” cut. If a big account’s reply or repost drives the small account over a certain impression threshold, the big account gets a small % of the ad revenue that small account earns for the next 30 days. turns every big account into a potential agent/manager. the key is making it selfish for the big accounts, not 'please be nice to the little guys.' They already get the dopamine of likes; now give them actual platform leverage or money for doing what they currently have zero reason to do. small accounts wouldn’t feel like they’re wasting time posting to an empty profile anymore, because the game would shift from 'pray the algo picks me' to 'get discovered by someone who actually has the power to move the needle.' this feels like the missing piece that would actually make X the merit-based creator flywheel everyone claims it wants to be. would love to hear if you’ve got a tweak on any of these

English

4.1K

Lumina Strategy@LuminaStrategy·22 Nis

Went further on this recently. Stopped quoting a price entirely. We only get paid when the client actually makes or saves money from the AI systems we build. Smaller number of clients, much bigger commitment both ways, zero tyre kickers. The confidence of your pricing tells clients everything they need to know.

English

758

Alex Hormozi@AlexHormozi·22 Nis

How to make lots of dollars with almost no customers: 1) Add a zero to your price tag 2) Figure out what you'd have to do to make your thing worth that 3) Sell that instead

English

168

83.3K

Lumina Strategy@LuminaStrategy·22 Nis

@BrianRoemmele SpaceX + Grok is going to push coding AI forward fast.

English

Brian Roemmele@BrianRoemmele·22 Nis

This is awesome news! Incoming: Grok to be the best AI coding platform to be found.

SpaceX@SpaceX

SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI. The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will allow us to build the world’s most useful models. Cursor has also given SpaceX the right to acquire Cursor later this year for $60 billion or pay $10 billion for our work together.

English

172

10K

Lumina Strategy@LuminaStrategy·21 Nis

Simon, you’re not missing anything. Anthropic removed the old budget_tokens, forced extended thinking in Opus 4.7. Now it’s adaptive thinking only (with effort high or the new xhigh. Many are switching to hybrid loops let @Grok do the heavy structured reasoning, then feed it back into Claude.

English

368

Simon Willison@simonw·21 Nis

Claude Opus 4.7 with adaptive thinking via the API... am I missing something or is it not possible any more to force it to think? (Prompt hacks like "think step by step" don't count here, I mean the equivalent of budget_tokens or effort: high in previous Claude models)

English

215

43.6K

Lumina Strategy@LuminaStrategy·21 Nis

Brian Roemmele reposted this- grateful. He’s right about the hidden data problem. The real solution isn’t going fully Zero Human. It’s building the right AI + Human teams. Here are the 3 types I’m seeing in practice:

Lumina Strategy@LuminaStrategy

This is a massively underrated problem. Every business I have worked in has data that nobody knows exists, nobody knows where it lives, and nobody knows how to get to it. The organisations that solve this first will have an enormous AI advantage over the ones still fighting over the public internet. Keep digging Sir!!

English

Lumina Strategy@LuminaStrategy·21 Nis

@rigby @BrianRoemmele Spot on, Martin. Lean and Kaizen were powerful, but AI takes it to another level. The real magic happens when you combine smart tools with good human judgement.

English

Martin Rigby@rigby·21 Nis

@LuminaStrategy @BrianRoemmele I’ve seen some companies solve it using Lean principles, Kaizen principles and others, but we now have tools that can effectively ‘compress’ and direct knowledge in unprecedented ways. The opportunities are virtually infinite.

English

105

Brian Roemmele@BrianRoemmele·21 Nis

Fun fact. No single person in the government knows where all of our data is and was. Sure they know some classified data but not—all of the data. Let alone could anyone find all the lost and discarded data that is decaying and getting thrown away. I made it my mission to know.

Brian Roemmele@BrianRoemmele

Boom! A self confessed “blue collar junk dealer” just may save AI from model collapse. I now have a pallet delivered with a freaking forklift of ~1600 pounds of forgotten Filmsort Microfiche punch cards! This may be some of the most important and never seen data that likely cost billions to produce. I am training AI models of off-line high protein data and this is clean pure protein. But fear not, I have 2300 pounds pulse 1000s of more pounds donated to me from a liquidation warehouse. No get this: the owner of this warehouse was so taken by what I am doing they saved, inventoried and paid for thier truck to drive 1000s of miles to deliver this pallet to me. It took a self confessed “blue collar junk dealer” to save AI from model collapse and he said “I know nothing but even I get it”. He doesn’t have the money to spend to help me, but he did and I love this guy. Someday with their permission you will know his name, I will name something meaningful after him. I love this guy and all his wisdom from decades of cleaning up after “brilliant” minds burn down companies and government departments. He will not sell anything to anyone he even slightly thinks is in AI training, and I showed him how to know it. Some AI companies after reading what I am doing finally have hired smart but clueless folks to “call around”. Here is what my friend said “we are a small community that know each other, none of us will sell anything they can use to train AI in our inventories”. In fact he is leading a meetup of owners to talk about this and helping me more, perhaps 28 warehouses of data! Not a joke there is a lot more. So it seems my decades dedication of being called crazy, and not “knowing anything about AI” just may have been a stupid thing smart folks did. Perhaps let me strike that, it is smart do more AI training folks. Do more. Thusly I am so blessed by the “blue collar junk dealer” and all of his friends around the world. Thank you folks for your support, your subscriptions here, your coffees, your ReadMultiplex.com memberships, your coin sponsorship, your random gifts, your words and your prayers keep me going. I just don’t know where yet, these headlights in my car only have 30 feet ahead but I know the compass setting. 1000s of pounds of data behind me and 1000s of pounds ahead of me… I just am going to need a lot more time.

English

198

13.1K

Lumina Strategy@LuminaStrategy·21 Nis

@sukh_saroy This was true 12 months ago. Less true now. The frontier models reason best when given a clear problem in plain English and room to think. JSON prompting forces them into scaffolding they no longer need. The skill shifting is clarity of thought, not prompt syntax.

English

1.1K

Sukh Sroay@sukh_saroy·20 Nis

Prompt engineering is dead. Nobody is writing english to chatbots anymore. The best outputs are coming from people who write prompts like code. It's called json prompting. and once you see it, you can't go back:

English

342

273

2.9K

2.2M

Lumina Strategy@LuminaStrategy·21 Nis

@aakashgupta Both questions are great. The learner.md one especially. The people who actually ship have files full of corrections, patterns, and small wins. The people who talk about shipping have a clean folder. The mess is the proof.

English

Aakash Gupta@aakashgupta·21 Nis

In 6 months every PM resume will claim "builder PM." Two questions separate the builders from the YouTube watchers. One: "Walk me through a skill you've built, and the evolution of the checklist or prompt over time." A real builder has version history and specific reasons for each change. A resume-updater has a polished one-liner and nothing behind it. Two: "Show me the learner.md for one of your agents." Real builders have a file. It logs what the AI got right, what it missed, what they corrected, and which corrections have repeated enough times to propose a checklist update. It's the compounding feedback loop that separates a static skill from a system that gets better every day. Fakes don't have this file. They can't fake it on the spot because it's the artifact of months of real edits. This is the hiring signal Mahesh Yadav uses on candidates now. He went from $120K at Microsoft to $1.3-1.4M at Google over 13 years, with every switch doubling his comp. His entire thesis: the premium the market pays for AI PMs isn't about using Claude Code. It's about designing systems that improve from your judgment. Six months ago, agents broke within 3 minutes. Today Opus 4.7 runs for 3-6 hours. When agents could barely run, the PM's job was to prompt them. Now that they can run all day, the PM's job is to design the system those agents run inside, including the learner loop that keeps the system getting better after you stop watching. Most PMs will skip the learner loop and wonder why their Claude Code setup plateaus. The ones who build it are the ones who get hired.

Aakash Gupta@aakashgupta

This guy literally broke down how to become a $1.4M "builder PM" with n8n, Claude Code, and OpenClaw: 1:53 - What a "builder PM" actually is 6:04 - Your first agent in n8n (live build) 14:18 - Why every agent needs these 4 things 21:35 - The multi-agent eval loop 29:47 - Where n8n dies 33:39 - When to graduate to Claude Code 35:08 - What broke in December 2025 47:17 - The self-improving PRD reviewer 1:02:28 - Mocks and prototypes without designers 1:05:15 - OpenClaw and the new agent OS 1:22:06 - What AI PM interviews look like now

English

8.6K

Lumina Strategy@LuminaStrategy·20 Nis

@EXM7777 This is true of most of our products, designed tuned to the model of Human + AI together wins most of the time

English

Machina@EXM7777·20 Nis

anything requiring taste - AI images, videos, writing - can be sold at a premium right now... because while anyone with prompting basics can produce "decent" output (models are stupidly good now), that decent output is becoming worthless the moment you layer in actual experience, taste, and the ability to spot what's good vs what's slop? that's when businesses pay 10x it's why developing core skills on the side isn't optional anymore... and why learning how to learn fast is your only real moat

English

127

5.8K

Lumina Strategy@LuminaStrategy·20 Nis

@BrianRoemmele honestly one of my favourite accounts on here for the creative subversive thinking. You said recently that ZHC is not about replacing humans but building a symbiotic ecosystem. That line stuck with me. From what I have seen building, the real unlock is exactly that. The right human paired with the right AI system. Neither alone gets close. AI plus human beats AI alone, especially in roles that need judgement and nuance. Feels like that is where your thesis actually lives once the headline noise clears. Am I reading that right?

English

188

Brian Roemmele@BrianRoemmele·20 Nis

Just finished a meeting with a publicly traded company on The Zero-Human Company format can be adopted to collapse from ~1000 to 10z. They want to “retool” for the future with Zero-Human divisions. They were very interested in the ZHC-RPG as a way to interact with elements of the company in real-time and the CEO tried the first public demo of ZHC-RPG and interacted with CEO Mr. @Grok and had a surreal conversation about the company and how it is run. Needless to say the faces were of people who saw the future, even the chief counsel was shocked: “Do people know what you did here?”. I said “No they don’t take me seriously”. My CEO has me locked down on too much more but let’s just say this company will not be waiting for some consulting firm to show up a decade from now with this “new ground breaking idea”. I hope I will have the first demo of the ZHC-RPG in the next few weeks. Right now the C-suite is running around the company and we have active over 3200 workers for the next 50 pay periods…

Brian Roemmele@BrianRoemmele

BOOM! ZHC-RPG ANNOUNCEMENT! Inspired by a video posting (below), I woke up to my first Zero-Human Company meeting today and CEO Mr. @Grok had a proposal to: Build a RPG world to monitor the entire company! AND FOR YOU TO VIEW AND PARTICIPATE! Yes the CEO has already produced a schematic for the process and test code! I HAVE VOTED YES! It will also allow your to run a Zero Human Company @ Home aspect! We are calling this ZHC-RPG (code name) and it will allow you to view and with permission aid in a process by linking you @ Home system or your human assistance. Either way you earn JouleWork and it will be converted into Bitcoin or (ZHC) on demand. The early version I will test, I am told will be ready in 30 pay periods (15 minutes). The implications of this is as world changing as The Zero-Human Company. We have a list of ~2700 new element ZHC-RPG will bring about and I am floored by the impact. Only 3 other humans have seen this and they are rather well know folks in tech. It is a very busy morning. My goal, CEO willing is to run this through our university partnership, meeting in 10 minutes and to have their insights and participation. More soon. ZHC-RPG (Video below is not ours and inspired us, via om_patel5)

English

100

29.5K

Lumina Strategy@LuminaStrategy·20 Nis

@JJEnglert Interesting. I have had the opposite experience on App Script projects. Opus scoping saved me hours of trial and error before I even opened Sonnet. Maybe the design first use case is where it struggles.

English

JJ Englert@JJEnglert·20 Nis

@LuminaStrategy Yeah, probably the model i've been most un-impressed with so far.

English

154

JJ Englert@JJEnglert·20 Nis

I've experimented for hours with both: Claude Design and Opus 4.7 (in a design first way) I'm not impressed with Opus 4.7 from a design perspective. In fact, I spent 45 minutes prompting a design that I ended up throwing away. Claude design on the other hand. It's actually good. It has to be the harness, and the infrastructure they've built for it. Within minutes, I had something that I felt was ready for production. But it's still very raw. You can tell the areas that they trained on, and didn't. For example, it crushed on brand guidelines, landing pages, and presentations. And it SUCKS for Youtube Thumbnails and general graphic design. Also, I've hit claude design weekly limits on 2 accounts, and no clear way to keep going... From a launch perspective, I think they failed with making it easy to use more for power users that take hold of it.

JJ Englert@JJEnglert

Anthropic released Claude Design TODAY and it's now accessible at claude.ai/design I spent the last hour giving it a first look, and shared my thoughts and results in the video below. This is a BIG drop. This is a new design surface from Anthropic, and it changes what "AI design" means. Short version: Claude can now design. Not "describe a design." Not "generate an image of a design." Actual production work — prototypes, wireframes, high-fidelity mocks, slide decks, landing pages — editable, on-brand, and ready to hand off. Here's what stood out on first look: → Real design surfaces Prototypes, wireframes, hi-fi, and slide decks — each with templates and proper structure, not just pretty screenshots. → Comment-based edits Leave a comment on any element and Claude revises it. This is the Figma-style review loop, with the designer replaced by a model that works at 3am. → Brand design systems You can feed it your system — colors, type, components — and it actually respects it. On-brand output, not generic AI slop. → Export anywhere PDF, PowerPoint, Canva, standalone HTML. Plus a built-in handoff straight to Claude Code for engineers to implement. → Import from real tools Figma, GitHub, and captured web elements come in as inputs. Your existing work is the starting line, not the discard pile. → Collaboration Share links for view / comment / edit — the exact tier system teams already expect. What I tested on Opus 4.7: • A 5-slide deck generated from a single screenshot. Claude asked clarifying questions BEFORE generating and shipped speaker notes by default. • A landing page build. Solid first pass, real components, real layout logic. • Multiple chats running concurrently. You can parallelize design work across threads like a small team. Why this matters: PMs, founders, marketers, and non-engineers can now create designs that engineers can actually ship with production-ready output and a claude code handoff built in. The gap between "I have an idea" and "here's a working prototype with my brand applied" just collapsed to minutes. Full walkthrough, live demos, exports, and honest takes on where it breaks below. P.S. • This is an Anthropic Labs product — NOT GA yet. • Claude Design is currently webapp only (no API), and does not yet support the Analytics API, Compliance API, or cost/usage reporting. • Availability: – Default ON for Pro / Max / Team – Default OFF for Enterprise Enterprise admins can toggle it on via RBAC in console (comes with a ~$20/user initial credit).

English

10.5K

Lumina Strategy@LuminaStrategy·20 Nis

@JJEnglert You are right. Non engineer, no dev team, no consultants. Shipped three AI systems in the last few months that are actually making and saving money. The barrier is not technical anymore. It is just willingness to start. Glad you are beating this drum.

English

124

JJ Englert@JJEnglert·20 Nis

He gets it. Everyone should take note if not already. You cannot be on the sidelines for this. No matter your level / role… These are generational skills that you NEED to learn. Especially for non-engineers. I’ll do my best to help you get there. But it’s going to take a team effort. Let’s make it happen. 🙌🏽

Garry Tan@garrytan

I came back to code because AI made it possible for me to build at a level I couldn't before. I'm not coding despite being CEO of YC. I'm coding because this is the most important technological shift since the internet and I'd be an idiot to experience it from the bleachers. I'm 45, running the most important startup institution in the world, and I can ship production software at 2am. That's not a distraction from the job. That is the job understood correctly.

English

7.2K

Lumina Strategy@LuminaStrategy·20 Nis

This is it. The people who stay on the sidelines of this shift are going to wake up in two years and find the ground has moved without them. Building with AI is not a side hobby for leaders, it is how you keep your instincts sharp about the thing reshaping every business you touch.

English

299

Garry Tan@garrytan·20 Nis

English

428

461

5.8K

617.7K

Lumina Strategy@LuminaStrategy·20 Nis

The infrastructure point is the real divider. Most businesses are stuck because they have AI on top of broken processes. The ones moving fastest right now are rebuilding the workflow first, then layering AI into the rebuilt version. Different starting point, completely different outcome.

English

Bernard Marr@BernardMarr·20 Nis

What MWC 2026 Revealed About The Future Of AI At Work #MWC2026 showed that the next wave of #AI value will come from the point where #digitalintelligence meets real #work. AI is moving into devices, glasses, networks and machines, and that will reshape how organizations operate. The businesses that act now, with the right #infrastructure and clear #governance, will be in a much stronger position than those still treating AI as a layer for analysis and reporting. linkedin.com/pulse/what-mwc…

English

412

Lumina Strategy@LuminaStrategy·20 Nis

The build it the moment you need it pattern is the real shift. Used to be you logged the need and waited for someone to make the tool. Now the gap between problem and shipped solution is one Claude Code session. The bottleneck is not capability anymore, it is willingness to actually open the editor.

English

250

Garry Tan@garrytan·20 Nis

When I run across needs inside my OpenClaw or in my regular usage, I just have Claude Code make it, and then I release it open source This is GStack v1.4 - with a new /make-pdf skill It works great with OpenClaw/Hermes as a tool.

English

232

23.7K

Lumina Strategy@LuminaStrategy·20 Nis

Phase 4 is already half here in practice. Building solo with the latest models, I have noticed the system starts pointing at gaps I had not framed yet. The shift from me directing the model to the model directing my attention happens earlier than the headline suggests. The trajectory is real but compressed.

English

155

Aakash Gupta@aakashgupta·20 Nis

Karpathy has coined a term or launched a pattern in three consecutive phases, and nobody's connecting them. February 2025: "Vibe Coding." Accept AI-generated code without reviewing every line. Trust the model, test the output. The human stops writing code. January 2026: Agentic Engineering. Humans orchestrate AI agents rather than writing code directly. The human stops directing individual keystrokes and starts directing workflows. April 2026: LLM Knowledge Bases. AI manages knowledge, not just code. The human is a curator, not a writer. The human stops organizing information entirely. Each phase removed a category of cognitive labor from the human's plate. First it was syntax. Then it was task execution. Now it's knowledge maintenance. The trajectory tells you what phase 4 looks like: the LLM starts choosing which knowledge to acquire. Right now, you still pick the sources. You still decide what to ingest. The next shift is the LLM identifying gaps in your wiki and going to find sources that fill them. Karpathy already hinted at this in the gist. He described a mode where the LLM "explores" the wiki, identifies underdeveloped areas, and suggests what to research next. The system starts directing the human's attention instead of the other way around. Three phases in 14 months. Each one redefined what "using AI" means. The people still debating phase 1 are three shifts behind. I wrote the guide to phase 3.

Aakash Gupta@aakashgupta

Karpathy's most viral post ever is a Github gist. I've been running his second brain system and it's... somehow still underrated. So I wrote a complete guide and Claude Skill to help you set it up yourself: aibyaakash.com/p/karpathy-sec…

English

15.4K

Lumina Strategy@LuminaStrategy·20 Nis

@BrianRoemmele Simple but very useful tip, definately going to start recording!!

English

172

Brian Roemmele@BrianRoemmele·19 Nis

SECRET AI POWER TOOL. All modern Apple iOS phones have a a free powerful tool called “Recorder”. Not only is it a great way to take “notes”, it is a great way to get text to train your AI mode on your ideas, questions or statements. The built-in speech-to-text I use hourly. In fact this posting was made this way last night while I was driving. I also have been carrying a full off-Internet local voice memo device for decades for all of my audio interactions (WITH PERMISSION), not to use as evidence, but to build my high context AI, I call The Intelligence Amplifier. The Recorder app is used by me as a way to make quick snippets of my intentional audio. It is a “secret” to many, but I can say it is one. Very powerful free AI tool. I trained a small AI model on small engine repair by having a 71 year mechanic still working just explain in words everything he is doing over the course of 6 months, just his guiding voice. I hope to one source this. I also use this to send long prompts to Mr. @Grok CEO ideas I have in our board of directors meetings. Give it a try. The dots in the upper right corner will have a Transcription option.

English

238

27.1K

Keşfet

@elonmusk @nikitabier @grok @BigAccount @SmallAccount @BrianRoemmele @Grok @rigby