πŸš€ (@quantumspaceup) header
πŸš€ (@quantumspaceup) profile picture

Founder @ That’s Gonna Help | Scaling πŸ—οΈ Growth πŸ“ˆ Revenue πŸ’° | Fractional CMO w/ Numbers Addiction | Speaker 🎀 | FinTech / Web3 / iGaming / SaaS

Web3
Joined June 2019
an anthropic researcher says claude helped him find a proof of the riemann hypothesis and published lean code. lean checks the proof, not whether the formal statement says what the headline says. independent expert review is still in progress. verdict after that. @AnthropicAI #AI #math
22
Most influencer deals get signed before anyone computes payback. The check I run first: 1. Take views, not subscribers. Subscribers are history, views are what the next integration can actually reach. 2. Estimate the paying share from your own past integrations, not from hope. 3. Adjust for geo. If you only monetize in some countries, count only that part of the audience. 4. CAC = integration price divided by expected payers. 5. Payback in months = CAC divided by monthly revenue per payer. 6. Write the formula down before the call, so the price conversation doesn't bend the math. 7. Set a threshold in advance. In my experience, around 3 months is normal. On a viral trend it can drop to weeks. 8. After launch, recompute with real payers and update the share for the next deal. Takeaway: a blogger isn't expensive or cheap. The payback number is. #marketing #AI
2
1
24
Replying to @GoogleDeepMind
@GoogleDeepMind says AlphaFold has predicted structures for hundreds of millions of proteins. a map that big moves the bottleneck to the wet lab that has to test it. the map isn't the territory. #AI #biology
11
nuclear engineer by training, so my bias: a model that fits the data hasn't explained it. AI makes fitting nearly free, so the discipline now is asking what result would prove the model wrong. #AI #science
5
a result announced by an AI lab and a result checked by outside experts are two different products. @AnthropicAI's CRISPR-like enzyme claim split researchers into excited and skeptical. wait for replication before quoting the headline. #AI #biology
when @OpenAI says its models proved a conjecture, the news is not the proof. it's that checking it now takes longer than finding it. verification is the scarce resource, and it still runs on humans with whiteboards. #AI #math
1
1
7
An AI creative factory costs less than one agency brief. How it works: 1. Start from winners. Pull your best performers and split them into parts: hook, visual, offer, call to action. 2. Vary one part at a time. Changing everything gives you noise, changing one thing gives you a lesson. 3. Approve stills before video. Fixing a frame is cheap, fixing a rendered clip is not. 4. Rent compute by the hour. Spin the GPU up for a batch, shut it down, pay for the batch and not for the idle time. 5. Add an automated pass for broken frames and artifacts, then a human glance at what survives. 6. Put the creative ID in the file name, so it carries into the ad platform and the report. 7. Launch in small batches and kill by a rule written beforehand, not by mood. 8. Feed results back. Winners become the seed of the next batch. Takeaway: AI makes variants nearly free. The scarce part is the rule for what gets killed. #marketing #AI
3
Replying to @OpenAI
@OpenAI described its safety approach from early research to internal use to public deployment. internal use is the stage outsiders see least, and my guess is where agents get real permissions first. publish incident counts from it. #AI #AISafety
2
Replying to @GoogleDeepMind
@GoogleDeepMind says gemini 4 neon sits between carbon and hydrogen and is aimed at science work. the mid tier is what labs will run daily, so ask for task-level results on messy lab data, not a leaderboard rank. #AI #AIforScience
5
Replying to @AnthropicAI
@AnthropicAI's ipo remarks from dario open with biology. the october claim of a claude-led gene-editing mechanism still has skeptics outside the lab. a listing raises the bar for replication, not lowers it. not advice. #ai #science
2
A branching video landing converted about 3-5x better than our best regular pages. How it is built: 1. Start from a page that already converted. Don't invent the format, take the structure that won and rebuild it as a story. 2. Break it into nodes. One opening scene, one choice, two or three branches, one peak. 3. Put the choice early. The viewer picks within seconds, and a picked path feels like theirs. 4. Gate signup at the peak, not before. By then they already want the next scene. 5. Generate the node videos with AI, but approve stills first. Fixing a frame is cheap, fixing a rendered clip is not. 6. Keep the clips light and host them on your own storage, so the page loads before attention drops. 7. Publish as a draft, test it yourself on a phone, then go live. Translate only after the original works. 8. Measure two numbers separately: visit to signup and visit to purchase. About a quarter of visitors signed up, and purchases ran several times above our best regular pages in the April audit. 9. Say the caveat out loud: fewer than 600 visits and 15 purchases in May 2026. A small sample, so treat it as a direction, not a law. Takeaway: AI makes the video cheap, the structure is what converts. Test it before you scale it. Follow for more operator notes. #marketing #ai
1
1
5
Replying to @GoogleDeepMind
@GoogleDeepMind's koray kavukcuoglu says no single benchmark will tell us we reached agi, progress is models getting less spiky. for operators that is the whole metric: the worst task in the workflow, not the best demo. #ai #agents
1
Replying to @OpenAI
@OpenAI says its models found up to 14% PPA gains on amd gpu ip in under five months, and gains on its own jalapeΓ±o inference chip verified on real silicon. "up to" is the best block. ask for the median across blocks. #ai #chips
Replying to @OpenAI
@OpenAI's research report: researcher ai use keeps rising exponentially, agents finish longer tasks, humans stay central. useful next ask: publish the denominator too. how many agent-hours end in a result a human keeps? #ai #research
5
Replying to @OpenAI
@OpenAI released gpt-6.2 astra and sol, 'super efficient and built for the real world'. that's launch copy. the useful part starts when independent evals land: same tasks, cost counted, someone else's grader. #AI #LLM
3
Outreach for a Bolivian tech week: 200 contacts, 3 closed. This is the funnel I ran. The setting: a new event, no sponsor list, and the outreach done by hand. Here is the process, as I run it. 1. Make the first message a short question. No pitch, no deck, no attachment. The goal of message one is a reply, not a sale. 2. Send one at a time, at a human pace. A few minutes between messages. Bulk sending gets accounts restricted and reads as bulk. 3. Use the channel where the person actually answers. I ran WhatsApp, email and LinkedIn side by side and let the replies show which one worked for whom. 4. Plan the follow-ups before you send the first message. A set rhythm over the first week, and every follow-up adds a new reason to answer, not a reminder. 5. Keep messages short. A couple of lines. If it needs a scroll, nobody replies. 6. Log every step in one CRM, the moment it happens. Sent, replied, refused, call booked. If it's not in the sheet, it didn't happen. 7. Count refusals and "next year" as answers. They are data about your offer and your timing. 8. Move interested people to a call fast. Threads explain, calls close. What it gave: 200 contacts, 30+ interested, 7 calls, 3 closed. The lesson: outreach is a funnel with a person at every step. Fix the step that leaks, not the whole thing. If you sell to companies, follow along. I share what broke and how I fixed it. #Outreach #B2BSales
4