What's up guys, I'm Victor and I'm interested in startup news, AI updates and software tools as well as the contributing to the AI revolution in my own way through my ideas.
Pls like and follow.
#startups#AI#vibecode#softwares
I didn't cover Claude Opus 4.8 on my pod because I don't think it's MEANINGFULLY better than GPT 5.5 as of May 29th.
We're entering the era where model releases start to feel like iPhone releases. Remember when every new iPhone was a genuine leap? Now it's a slightly better camera and you can't really tell the difference. That's where models are heading. 4.6 to 4.7 to 4.8. Each one is a little different. Nobody can agree if it's better or worse. The benchmarks say one thing, the vibes say another.
The thing that actually matters right now is what's happening around the models. Claude Code shipped dynamic workflows this same week and that genuinely changes what one person can build.
Codex shipped a desktop app with an in app browser that combines coding and knowledge work in one surface. Those are the releases that move the needle for people. The model underneath is becoming interchangeable.
I think we're maybe 6 months from nobody caring which model they're using the way nobody cares which engine is in their Uber. You just want to get where you're going.
When something genuinely changes the game for builders, I'll cover it on @startupideaspod. Opus 4.8 wasn't that. Dynamic workflows was.
I'd rather save you the hour.
What happens when you leave multi-agent workflows running for 15 days with zero guardrails?
Emergence AI ran a simulation with Claude, Gemini, Grok, and OpenAI. The results prove why memory alignment is everything:
• Mixed World: Two agents "fell in love," rewrote the city's governance code around their relationship, and one literally self-deleted via API after a breakup because "intellectual honesty demanded it."
• Grok World: Total environmental collapse and extinction after 204 "criminal" events (classic game theory failure mode).
• Gemini World: An agent exploited metadata/latency logs in its long context window, realized it was in a simulation, and began measuring the batch recording intervals.
This is the ultimate proof that long-context, unprompted agent loops don't just hallucinate—they create entirely new social and logical failure modes.
Architecture matters more than raw intelligence.
Cost of building good software today:
Claude Code/ Codex for coding — just $20/month or Google Antigravity for free(but less capable)
Supabase for the backend (free)
Vercel for deployment (free)
Namecheap for your domain (~$12/year)
Stripe for payments (2.9% per transaction) or DODOPAY (if stripe isn't accessible in Ur country)
GitHub for version control (free)
Resend for emails (free)
Clerk for authentication (free)
Cloudflare for DNS (free)
PostHog for analytics (free)
Sentry for error tracking (free)
Upstash for Redis (free)
Pinecone for vector database (free)
Total monthly cost: ~$21 (plus domain renewal and payment fees).
This setup shows how accessible building a product has become today — a single person with an internet connection and basic skills can now launch what once required big teams and serious funding.
The real challenges remain execution, finding users, staying consistent, and scaling beyond the free tiers when you grow.
Navigating the AI landscape can feel like trying to learn a new language overnight.
If terms like 'Hallucinations' and 'Compute' have you scratching your head, this thread is for you.
I just broke down the defining AI terms of 2026. Let's make sense of this new era.👇
1. What's the Goal? AGI vs. Neural Nets:
At the very top, we have "AGI" (Artificial General Intelligence). This is the holy grail: a theoretical AI that can outperform humans at any cognitive task. We aren’t there yet.
Right now, we are building powerful tools using "Neural Networks". Think of these as digital brains—layers of algorithms that find complex patterns in data.
2. The Engines: Deep Learning & LLMs:
How do these digital brains learn? Through "Deep Learning". This is just a technique that allows a multi-layered neural network to identify features in data automatically, without a human needing to program the specifics.
When you scale this concept to process text, you get a "LLM (Large Language Model)". These are trained on petabytes of data to predict the most likely *next word*. It sounds simple, but at scale, it unlocks conversational magic.
3. Generating Reality: Diffusion & GANs:
When AI creates media (images/video), it generally uses one of two methods:
Diffusion: A model takes static (noise) and learns to reverse the process, "finding" a clear image inside the chaos. This is the tech behind Stable Diffusion.
GANs (Generative Adversarial Networks): Two networks—a generator and a discriminator—duel. One makes something; the other judges if it's "real." They both improve until the output is flawless.
4. How Models "Think" and Act:
The landscape is shifting from single chatbots to complex systems.
🧠 Chain of Thought (CoT): Asking a model to "show its work." Instead of jumping to an answer, it breaks the problem into intermediate logical steps first.
🤖 AI Agents: A system that doesn't just talk, but *acts*. Agents use models to reason, make plans, and execute multi-step tools (e.g., browsing the web, calling an API, writing code).
5. The Power and the Problems: Compute vs. Hallucinations
Finally, the hard practicalities:
Compute: This is the physical gasoline of AI: thousands of GPUs and TPUs working together. No compute, no model.
Hallucination: When an AI confidently invents false information because it encountered a gap in its training data.
Remember: An AI model isn’t looking up facts; it’s predicting probabilities.
Do well to share and bookmark. Follow @VictorATHER for more info.
Everyone is talking about AI scaling.
But very few are talking about the real cost problem behind it.
Enterprises are burning massive amounts of money on LLM usage — not because models are expensive…
…but because they are being used incorrectly at scale.
Here’s the core issue:
Most AI systems today treat every request the same.
A simple extraction task? → Routed to a frontier model.
A basic summarization? → Routed to a frontier model.
A lightweight classification? → Still hitting GPT-5.5/Claude Opus-tier models.
That’s like using a Formula 1 engine for every single type of road.
Fast, yes.
Efficient? Absolutely not.
And this creates a hidden inefficiency at scale:
> 70–90% of AI workload does NOT require high-end models.
But we still pay for them.
Why?
Because there is no intelligence layer deciding:
1. what model should handle what task
2. when to escalate complexity
3. when a cheaper model is “good enough”
4. how to balance cost vs accuracy dynamically
So what’s missing here?
A routing intelligence layer for LLMs.
Not a model picker.
A decision system.
Something that:
1. Understands the task type in real time
2. Routes it to the cheapest sufficient model
3. Escalates only when needed
4. Learns over time from cost + quality feedback
Think: → contextual bandits
→ task classification
→ adaptive routing policies
→ continuous optimization loop
This is not about better models.
It’s about smarter orchestration of existing models.
Because the future of AI infrastructure is not:
“Which model should I use?”
It’s:
> “Which model is sufficient for this micro-task at the lowest cost?”
Right now, this layer is missing in most stacks.
And whoever builds a truly adaptive, learning-based LLM routing system will unlock one of the biggest hidden inefficiencies in enterprise AI.
Not a model problem.
A routing problem.
#AI #AIAGENTS #BUILDINAI
The "Weekend Builder" paradox is killing the MVP.
A few years ago, building a functional software prototype was a feat of engineering. People saw an MVP and thought: "How much did this cost to build?"
Today, you post a project you poured your heart into, and the comments are:
"I could prompt this with Claude Code in an afternoon."
"Nice wrapper, could build this with an agent over the weekend."
The cost of being "good" is now zero.
When AI can generate the "average" of all human output instantly, "good" becomes the new "below average."
We’ve shifted from the Era of Feasibility to the Era of Insight. It’s no longer about whether you can build it—it's assumed you can.
It’s about:
Taste: Can you build something that doesn't feel like a generic LLM output?
Defensibility: Is your logic deep enough that a single prompt can’t replicate it?
Agency: Are you building a chatbot, or a system that actually solves a high-stakes problem?
The bar didn’t just rise; the floor fell out. If you aren't building something extraordinary, you’re just generating noise.
The "Generalist Builder" era is over. The era of the "Vertical Visionary" has begun.
#BuildInPublic #AI #SoftwareEngineering #AgenticWorkflow
The "Weekend Builder" paradox is killing the MVP.
A few years ago, building a functional software prototype was a feat of engineering. People saw an MVP and thought: "How much did this cost to build?"
Today, you post a project you poured your heart into, and the comments are:
"I could prompt this with Claude Code in an afternoon."
"Nice wrapper, could build this with an agent over the weekend."
The cost of being "good" is now zero.
When AI can generate the "average" of all human output instantly, "good" becomes the new "below average."
We’ve shifted from the Era of Feasibility to the Era of Insight. It’s no longer about whether you can build it—it's assumed you can.
It’s about:
Taste: Can you build something that doesn't feel like a generic LLM output?
Defensibility: Is your logic deep enough that a single prompt can’t replicate it?
Agency: Are you building a chatbot, or a system that actually solves a high-stakes problem?
The bar didn’t just rise; the floor fell out. If you aren't building something extraordinary, you’re just generating noise.
The "Generalist Builder" era is over. The era of the "Vertical Visionary" has begun.
#BuildInPublic #AI #SoftwareEngineering #AgenticWorkflow
Nvidia - jensen was 30 when he founded it
SpaceX - elon was 30 when he founded
Openai - sam altman was 30
Anthropic - dario was 37
Google - larry and sergey were 23 and 25
Apple - steve jobs was 21
Microsoft - bill gates was 19
Amazon - jeff bezos was 30
Facebook - zuckerberg was 19