Perplexity is losing users but making more money than ever. Make it make sense.
• Sessions down 38% from peak
• App installs down 93%
• CEO just announced $1 BILLION ARR
Everyone's calling it a downfall.
Meanwhile, the business is apparently booming.
3
4
234
Forget the hype. Which personal AI agent is ACTUALLY the best?
- Hermes Agent
- Grok @bot
- Muse
- ChatGPT Dots
- Claude Duet
- Openclaw
Or is there something better?
2
151
You believe in an idea. You launch a waitlist.
It completely flops.
Do you kill the idea or build it anyway?
If Brian Chesky had launched a waitlist for Airbnb before building it, would anyone have signed up?
Sometimes I wonder how many billion-dollar ideas would've died at the waitlist stage
2
3
151
The evolution of AI hallucination
2023: "Sorry, I made that up."
2024: "According to my knowledge..."
2025: "According to these sources..."
2026: "Your sources are wrong." 💀
AI hallucinations in 2026 are MORE DANGEROUS than they were in 2023.
Back then, AI would make up facts, names and sources. The mistakes were often obvious.
Now it's much harder to tell.
I asked Claude Sonnet 5.5 to analyze 5 research papers.
It returned detailed statistics, DOIs, citations, and references to specific tables.
Everything looked legitimate.
Then I asked it to audit it's own answer.
It admitted to making claims it hadn't verified, misrepresenting source accessibility, and misinterpreting findings.
And I probably wouldn't have caught those errors otherwise.
That's what scares me.
AI isn't just getting better at being right.
It's also getting better at sounding right when it's wrong.
2
91
Microsoft found a way to make AI coding agents up to 31% cheaper.
Their research uncovered 3 major problems:
- 30% of tool calls repeated previous work.
- 16 out of 23 available tools were never used.
- Failed tasks consumed 32% of total spending.
So they built TokenEconomizer.
It remembers repository-specific lessons, removes unnecessary tools, and intervenes when agents get stuck in failure loops.
The result:
18–31% lower costs across 277 coding tasks, with no meaningful drop in success rates.
The future of AI coding isn't just smarter models.
It's smarter agents.
2
5
172
AI hallucinations in 2026 are MORE DANGEROUS than they were in 2023.
Back then, AI would make up facts, names and sources. The mistakes were often obvious.
Now it's much harder to tell.
I asked Claude Sonnet 5.5 to analyze 5 research papers.
It returned detailed statistics, DOIs, citations, and references to specific tables.
Everything looked legitimate.
Then I asked it to audit it's own answer.
It admitted to making claims it hadn't verified, misrepresenting source accessibility, and misinterpreting findings.
And I probably wouldn't have caught those errors otherwise.
That's what scares me.
AI isn't just getting better at being right.
It's also getting better at sounding right when it's wrong.
- SpaceX acquired Cursor.
- Meta acquired Conductor.
- OpenAI acquired Astral, Ona, and Pierre.
- Anthropic acquired Bun.
- Cognition acquired Windsurf
Something massive is happening in the AI coding industry
3
3
338
Developers who've never worked at tech companies may never realize how important understanding your code actually is.
On personal projects, "it works, don't touch it" can survive.
In production, you'll have to touch it sooner or later.
Requirements change, bugs happen.
The more you don't understand your code, the harder things get.
1
112
Claude Opus 5.5 is becoming my go-to model for software architecture.
Even over GPT-6 Astra and GPT-6.1 Sol.
Here's what stood out while designing my AI coding agent:
• Separation of concerns: Opus pushed for clear component boundaries instead of one giant orchestrator doing everything.
• Failure handling: It considered what happens when models fail, tools time out, or subagents return conflicting results.
• Trade-offs: It didn't just suggest components. It challenged whether I needed them, explained the complexity they introduced, and when they'd actually be worth adding.
The difference wasn't the architecture diagram.
It was how deeply it thought through the decisions behind it.
That's what I want from an AI when designing a system I'll actually have to maintain.
8
15
652
I spent 12 hours reverse engineering how AI coding agents work.
Here's what surprised me the most:
1. It's literally just a while loop. Call the LLM → execute its tool requests → feed results back → repeat.
That's the core of Claude Code, Codex, and Cursor.
2. The model has amnesia. Every API call starts fresh. The harness has to supply the conversation history again. The LLM isn't actually remembering anything.
3. The AI never touches your computer. It just requests actions. A regular program called the harness reads files, executes commands, and writes code on its behalf.
4. The model is only half the story. Give two coding agents the exact same LLM, and they can perform wildly differently. The difference is their tools, context management, and orchestration.
The magic isn't just in the model. It's in the system built around it.
I'm building an AI coding agent with UNLIMITED usage on frontier models like Claude Opus, Sonnet, and GPT-6
And I'm starting to understand why this is so difficult.
Ever since I announced artesian, everyone keeps asking:
- "Will you host local models?"
- "Which open-weight models are you using?"
- "Is it BYOK?"
Everyone seems to assume unlimited usage means settling for cheaper less capable models.
But that's not what I'm trying to build.
What's the point of unlimited coding if the agent can't reliably get the job done?
I want artesian to compete on capability AND reliability, without the frustrating usage limits.
And yes, that means expensive frontier models.
That's what makes this problem so much harder.
Will there be trade-offs? ABSOLUTELY!
But those trade-offs will NEVER be quality or reliability.
Can I actually pull this off?
I genuinely don't know.
But I KNOW it's worth finding out.
If you believe this is worth building, join the artesian waitlist.👇
Every signup gives us another reason to keep pushing.
3
6
308
I wish I knew these 7 Claude Code features SOONER:
1. Claude can't quit until your tests pass.
A Stop hook blocks it from finishing. It keeps fixing until it's green
2. Claude broke the code? Press Esc twice.
Restore the code, the conversation, or both.
3. Your phone can run it.
/remote-control let's you control a session on your laptop from anywhere
4. It works while your laptop is off.
/schedule runs tasks in the cloud, on a timer or a GitHub event.
5. /btw asks a side question without polluting your context.
Your main task stays clean
6. Auto mode: a second AI reviews each action instead of you.
It blocks the risky ones.
7. Subagents can run different models.
Haiku for quick reads, Opus for hard stuff. Same session.
Bookmark this. You'll need it🔖
9
9
545
Why are we spending billions teaching AI to do EVERYTHING when all we want is for it to CODE?
I couldn't stop thinking about this.
Why not build an AI trained exclusively for coding? Pour all that compute into making it the best programmer alive.
And whenever it needs real-world knowledge? Just let it consult a general-purpose LLM.
It sounded like a cheat code in my head (silly me)
So I went down the rabbit hole (a 37 minute convo with Fable 5.2)
And turns out, here's all I missed to consider:
Being great at writing code != being great at software engineering.
To build great software, AI needs to understand mathematics, human behavior, economics, security, and the actual problems it's solving.
So how can it know the right questions to ask the gp LLM if it doesn't understand these?
And general-purpose models are also heavily trained on coding.
My biggest realization:
The best coding AI isn't the one who knows the most code, it's the one who understands the problem best.
2
7
507
My biggest problem building artesian isn't writing the code.
It's figuring out how to offer unlimited AI coding without sacrificing frontier intelligence and reliability, while also not going bankrupt.
Every prompt costs money.
Heavy users can consume far more than they pay.
And I have basically no funding to subsidize them.
That's the problem I'm trying to engineer my way out of.
14
10
639
Sonnet 5.5 is the most underrated model in Claude's family right now.
Half the price of Opus 5.5
Same score on independent tests (AA-Briefcase: 1813 vs 1800).
30%+ Faster than Sonnet 5.
Fable and Opus get all the hype in the family but Sonnet gets the work done.
Opus and Fable still win on the hardest tasks.
But for everyday coding and agents? Start with Sonnet.
How to go from a vibe coder to AI-assisted programmer in 7 steps:
1. Plan first, code later
Prompt: "Don't write code yet. Give me a step by step plan and ask me questions".
Approve it before you let it write any code.
2. One small piece at a time
Never ask "build me the whole app." Build one feature, test it, then move on.
3. Make it explain itself
Prompt: "Explain what you just wrote like I'm a beginner. What could break?"
If you can't follow it, don't ship it.
4. Save every working version
Use Git. Commit after each thing that works. When AI breaks something, you can easily roll back like it never happened.
5. Ask how to test it
Prompt: "How do I verify this works? Give me all the edge cases that could break it. Write tests for it."
Don't just trust it because "it runs".
6. Get a second opinion
Open a new chat, paste the code.
Prompt: "Review this for bugs and security holes."
AI is a better reviewer when it didn't write the code.
7. Understand errors, don't just paste them
Ask "why did this happen?" before asking it to "fix it". Every bug teaches you a free lesson about the project.
Did I miss a step?
7
5
462






