Founder https://nitter.gq/t.co/zBq52j1B4q | Co-founder https://nitter.gq/t.co/ASMTyiNEWH
Joined September 2010
- Tweets3.1K
- Following94
- Followers198
- Likes12.4K
If I were CEO of Apple, I’d focus on making Apple hardware as open as possible for AI companies. Turn Apple into the hardware of choice for AI.
Apple already makes great hardware. You can shout “Hey Siri” from anywhere in the room and it listens. Now imagine: “Hey Siri, ask Claude to do this” and Siri just hands the request off to Claude (or any other model) running on the device. All developers would need to do is hook into the Siri APIs.
Apple doesn’t need to build its own models. It can become the foundation where these models run and enter people’s lives. There’s immense upside to it with very little CapEx. Kind of what DHH is doing with Omarchy for Linux.
Software is dead
Software is not dead
You will hear good arguments from both sides
But nobody knows shit for sure
Ignore the noise and build whatever the fuck you want
ppl keep asking me how u even measure roi w aeo?? like why does chatgpt even mentioning u matter? how does that actually bring more business??
heres how i explain it.. aeo is completely different from seo
before ai, seo was pretty much just google search
now aeo is google + chatgpt + claude + grok + whatever else
and these ais dont just dump blue links. they actually read ur stuff and form a sentiment/opinion
that changes everything. if they talk about u negatively and u do nothing, itll hurt ur business long term
think of it like a bunch of influencers talking shit about ur brand on insta. trust just shatters. so aeo is reputation management too
not just sentiment either. these platforms also pull from socials.. youtube transcripts, reddit threads, insta reels, all that. so aeo is becoming social media marketing as well
lot of ppl think aeo is just making a bunch of x vs y blogs. nah not even close. its way more than that. to actually win at aeo u gotta work on ur brand as a whole
Product/business ideas you can build with Jev by @typesafeai
This list is based on all the ideas I have seen here on X shared by other people.
1. Agent loops (the loudest theme)
- Pick the next tool / subagent
- Decide continue / retry / ask the user / stop
- Reject bad tool results before the next step
- Route “fast vs plan” or flash vs SOTA models
- Score which bots in a multi-agent group should take a task
2. Guardrails and “trust the agent more”
A popular idea: let agents act more freely if Jev scores each sensitive action for scam / prompt-injection / policy risk and blocks low-confidence moves. Same pattern for verifying LLM outputs, citations, tool calls, and jailbreaks cheaper than running another frontier model.
3. Support, tickets, and ops routing
- Classify tickets (sales / support / spam / billing / tech)
- Score urgency, frustration, refund risk, churn
- Pre/post-triage scoring on Jira/Linear and support queues
- Zapier-style lead routers replaced by Jev
- Auto vs human-review based on confidence
4. Content filters people are already shipping
- Classify X notes as LLM-written or engagement-farming and hide them
- Block YouTube sponsorships / hide ads / detect LLM text in real time
- Live debate scoring
- Self-evolving classifiers: dump “other” cases, have an LLM propose new categories, keep Jev as the fast classifier
- An “awesome-jev” list where Jev itself reviews submissions (evidence, description, category) before a maintainer merges
5. Security engineering
- Rank hunt results and score hosts/users/processes as suspicious
- Classify noisy events (e.g. rundll32) at scale
- FP analysis on historical alerts
- Incident response: shrink timelines, prioritize scope expansion
- Continuous compromise scoring
6. Search, routing, and product gates
- Search-intent gate: chat vs search / should we hit the web? (replace the extra flash-model hop)
- Local model gateway: cheap model vs Claude based on query type
- RAG: score query–doc relevance, rerank, pick context
- Semantic linting in CI (code + writing conventions)
7. Big-data / cheap map-reduce
Because input is ~$0.042/M tokens and output is free, people talk about judging everything:
- Score every document a company ever wrote (one demo: 37 docs × 21 questions in ~0.7s)
- Feature extraction for classical ML (intent, risk, churn signals)
- Pair Jev with extractors so humans only review low-confidence rows (founder mentioned this for giant enterprise piles)
8. Quirky / speculative
- Robot action selection: grasp / rotate / inspect as tools; tiny fast model picks the next primitive (compared with other “no-chat” models)
- Game / real-time UIs (official docs even mention games)
- “Ask Jev” as a product surface; jokes about central-bank decisions
- Open-source “Jev-like” projects on your own GPU
I have a feeling this will be a sleeper hit!
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Precisely why we went with Chat-first interface for AEO work at Zerply.ai and not complex drag and drop workflow builder UI. Talking is natural. Nobody wants to use more software.
New AEO hack just dropped.
Google launched an embeddable “Preferred Sources” button.
You drop a simple script and div on your site. Readers click it and mark you as one of their preferred sources. From that moment, your content becomes more likely to show up (with a clear Preferred badge) in their Top Stories, AI Overviews, and AI Mode.
Google’s own data already shows preferred-badged citations pulling roughly 2x the click-through of unlabeled ones in the same AI response. That’s not a small edge.
They also added natural language controls so users can tell Discover what they want more or less of, almost like texting a friend.
This is Google quietly giving users more control over what they see in AI search. And that control flows directly to the sites people already trust.
Most sites will treat this as optional. A few will make it dead simple for their readers. Those few will start showing up more often inside AI answers while everyone else keeps hoping the algorithm notices them.
We’ve already turned this on by default for every blog hosted on Foundry (zerply.ai/platform/foundry). No extra setup required.
If you publish online and care about showing up in AI search, this is one of the easiest levers available right now.
Claude and GPT are the default models for most of us. They are amazing, but they cost a lot per run. Using them for every blog draft, competitor check, and outline is how a week of ordinary AEO/SEO work gets expensive.
If you are doing your own marketing, you cannot afford the expensive model on every draft. You also cannot afford a dumb one either.
But lately, open-weight models like Kimi, Qwen, DeepSeek and GLM are giving head-on competition to models from OpenAI, Anthropic and Google, at a fraction of the cost.
So we added these cheaper, albeit amazing models to Zerply Agent. You can now use Kimi K3, Qwen, DeepSeek V4 Pro, GLM 5.3 (and now Grok 4.6 too) in Zerply chat. Open a new or existing chat, and pick the model you'd like to take for a spin.
When you start a new blog post or landing page in Zerply, that same model assortment is there too.
Speaking of benchmarks, Kimi K3 scores 57 on Artificial Analysis's Intelligence Index, a public ranking of how capable these models are. Claude Fable 5 scores 60. GPT-5.6 Sol scores 59. DeepSeek, GLM, and Qwen cost much less, so everyday keyword research and first drafts do not have to run on Claude or GPT.
And a friendly reminder: Zerply chat is just as capable as Claude or ChatGPT app, and replaces multiple subscriptions into one. Plus, you (and your team) get the benefit of using any model available today and in the future.
Your prompts and data stay on 🇺🇸 US infrastructure, fully private.
More info on:
1. Zerply Agents: zerply.ai/platform/ai-agents…
2. Signup for a free 7 day trial: app.zerply.ai/signup
Add this to your claude/cursor rules: “Respond only in ASD-STE100 Simplified Technical English”
Thank me later.
Preetesh retweeted
The limiting factor of software engineering has never been code production. It is complexity management.
Most people are using Opus 5 wrong. Give it a simple task and it will waste tokens and you will think its dumb. But give it a really big & difficult task, and it will not stop until it is done to the level of perfection you prompted it to.