Oz (@FoxOnTheRunTr) header
Oz (@FoxOnTheRunTr) profile picture

SaaS Solopreneur. Digital nomad. 🌏💻 1) MacOS File Mng https://xitter.cf/t.co/KTIQUtCq97 2) https://xitter.cf/t.co/2ZKEITy0WT 3) https://xitter.cf/t.co/yM5R5r6oWl Hustle is the name of the game

▀▄▀▄▀▄ 𝙃𝙐𝙎𝙏𝙇𝙀 & 𝙁𝙇𝙊𝙒
Joined February 2013
Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)Photo by Oz (@FoxOnTheRunTr)
"Hi Jev," Last week another AI model broke the internet. I spent the weekend on X (Twitter) learning about it from others and tried it on a few use cases. Here is a relatively long post on my learnings if you want to stay up to date on the latest in AI. The model is called "Jev" from a company called typesafe ai which came out of stealth on 15th September, co-founded by Diogo Almeida who spent ~4.5 years in OpenAI doing technical data science work. How is this model different? • It's not "conversational" model, it is a classifier. This means you throw some text at it along with a list of "tags", and the model returns which "tags" best apply to this text along with the probability / confidence. • It is super efficient at doing this because its architecture is entirely different to a traditional LLM. • A traditional LLM (Claude, OpenAI, etc.) architecture runs a so called "autoregressive decode loop" when it needs to give you an answer, which is expensive and it answers in most often unpredictable "text" rather than some binary (yes/no) classification). Jev is different, its output is fast and short. • Jev is ridiculously fast. It gives you an answer in 70-500ms (you barely notice it basically). Traditional smaller/faster LLM models give you an answer in 1-3s, and here we are talking about the small (not so "brainy" models), not the large brainy ones (Fable, Sol). So Jev is crazy fast compare to them. • Jev is also ridiculously cheap. Models are costed on price per million input and output tokens (1 token = 4 characters approximately). Jev only costs $0.042 per million input tokens, and output is free (which is not the case for any other LLM). For the sake of comparison, the cheapest models available from Google and Open AI cost $0.05 input and $0.2-$0.4 output. Why does this matter so much? • Because a significant portion of the work done by the world is pure classification. Industry studies show that around ~70%-85% of production machine learning (ML) models used by global enterprises are pure classification work, not "chat" we are used to seeing. • And if that's the case, then the unit economics are crazy. If you compare Jev to frontier models (Sol, Fable) then Jev is ~25-80 faster and ~200-600x cheaper. If you compare Jev to mid-tier models (Luna, Gemini Flash), then Jev is ~5-18x faster and ~6x cheaper. Hence the hype. What did I learn from testing it? • I signed up to the wait list for Jev last week when it came out, and soon after was given an account with $5 test credits. • I tested using Jev for automatically tagging conversations from Intercom, and everything above is true. The speed, the price. It's all there. I have the feeling that even running out of $5 is difficult. When I did similar tagging via Gemini Flash Lite, I burned through $5 quite quickly. So what do I think will happen now? • Jev is here to stay and will be used by many. The hype is real. • Jev will be integrated into many production systems and replace traditional LLMs there when it comes to classification work (I am already doing this on some personal projects). • Jev will enable better user experience. You can now do almost real time classification / decision work, so you can do some very interesting experiences. Remember: classification is just a decision. A decision can be to decide if email is spam or not spam, but also whether a character in a game should turn left or turn. That's a classification decision too. What are the top use cases I can think of? • High-throughput content moderation and safety gating, e.g. social feeds, gaming lobbies, etc. Jev evaluates harassment, self-harm, scam solicitation and overall severity concurrently in a single 70–150ms pass. • Enterprise ticket triage and semantic routing. One call classifies department, assigns urgency Score, and flags churn/legal-escalation. • Autonomous agent tool selection and action dispatch. Jev as the routing brain. Sometimes when you chat with an AI system, the system needs to determine should it call a large "expensive" model (Fable) or a cheap model (Sonnet). Jev can route your request to the right model depending on complexity. • Execution guardrails, risk scoring and escalation gates. Inspecting proposed code, terminal commands or database operations pre-execution. A software engineering use case. • LLM output verification, fact gating and hallucination auditing. Jev as sub-second judge over a larger model's draft: does the draft contradict the retrieved context, are citations accurate, are required disclosures present. • Bulk data labelling. Generate synthetic labels over unannotated corpora, filter strictly to instances where confidence exceeds e.g. 0.95, and use those high-confidence. • Real-time reactive control loops: games and web automation. Sub-100ms enables genuine feedback control. There was a Doom game demo which showed Jev controlling the Doom character.
6
4
1
13
1,470
GPT-5.3 Codex absolutely demolished Opus 4.6 (65.4%) on Terminal Bench 2 just minutes after its launch!
39
52
8
731
66,467
OpenAI has now: 1. Signed $500 billion Stargate deal 2. Signed $100 billion Nvidia deal 3. Signed $100 billion AMD deal 4. Signed $38 billion Amazon deal 5. Signed $25 billion Intel deal 6. Signed $20 billion TSMC deal 7. Signed $13 billion Microsoft deal 8. Signed $10 Billion Oracle deal 9. Signed "Multi-Billion Dollar" Broadcom deal 10. Launched a browser to compete with Chrome 11. Become the world's most valuable private company 12. Considered $1 trillion IPO by 2027 We are in the midst of a generational technological revolution.
969
1,401
542
10,862
3,649,053
Your ability to code is your floor. Your ability to communicate is your ceiling. Most developers are building amazing basements with no way to the top floor.
1
1
10
158
Two founders: Founder A: "We use advanced ML algorithms to optimize user engagement through predictive analytics..." Founder B: "We help apps keep users coming back." Guess who is the winner? Simplicity is an important skill, both in communication and in design
1
1
4
326
The best technical solution rarely wins. The best communicated solution does. How are you packaging your genius?
2
1
8
310
Day 43/100 common indie dev pitfall to avoid 🚨
3
1
9
352
I'm sipping coffee and the terminal agent's helping me with project analysis. This is such a nice vibe! Thanks for the image optimizer app, by the way. @FoxOnTheRunTr
1
5
151
Launching soon! Please be there to support!! First Product Hunt Launch
2
2
1
91
Good news for the tech job market buried in the “big beautiful bill” - companies will be able to fully and immediately deduct software development costs (aka your salaries) again
118
375
137
5,371
709,129
ELEVENLABS WHO? This voice AI is open-source, ultra-expressive, and actually beats ElevenLabs in quality. No paywalls. No weird licensing. Just pure, controllable speech synthesis — built for creators and devs. Here’s the model everyone’s switching to: 👇
210
526
44
5,998
1,435,425
Just subscribed to Github Pro + Sentry issue -> Github issue -> Affected to Copilot -> PR automatically created with perfect solution This is insane 🤯🤯🤯
2
2
4
302
Sir Demis Hassabis is the most dangerous CEO alive: • Chess prodigy at age 4 • Knighted in 2023 for services to AI • Nobel Prize in Chemistry in 2024 He now leads Google's DeepMind AI His vision of the next 10 years will terrify you 🧵
298
1,548
163
13,155
3,098,707
Ever shared a screenshot and felt like... something’s missing? That’s what CrispShare fixes. Built by @FoxOnTheRunTr, it lets you turn any screenshot into a scroll-stopper — • X cover • Insta post • LinkedIn banner • Even X double-photo layouts I just made the attached image using it — the entire process felt like dragging pieces into a vibe. ✨ Add backgrounds (color, gradients, images) ✨ Customize layout, borders, curved corners ✨ Blur what you want ✨ Drop your watermark Oh, and yes — it runs smooth. I’ve seen tools like this before, but none this clean. Give it a try → CrispShare.com Shoutout again to @FoxOnTheRunTr — absolute game-changer.
Day 27 🌴: - switched to Gemini Flash 2.5 for the copilot of Form Builder project (Haiku doesnt have structured output like Gemini does), gonna test it out for a while - yesterday released a free software CrispShare to showcase screenshots like this. Your tests are most welcome!
1
2
178
Started vibe coding a graph compute engine kind of like n8n or comfyui over the last couple of days. My plan is to make it extremely easy to create/one-click deploy APIs I can use on my servers without having many external dependencies, but while having all the niceties of things like queueing, rate limiting, versioning, webhooks, etc. I'm not planning to make it a public tool, just a fun little side project that I hope could enhance my own turnaround times and maybe be able to share with less technical partners to do things like make ai agents. The node compute can be arbitrary, so there are not many limits. Shoutout to @FoxOnTheRunTr for the screenshot editing software (crispshare.com)
Working on a n8n alternative for my own use case. Going very well so far.
1
1
1
3
231
so beautiful....AGI soon?
i asked gpt4.5 what’s a deep truth that most people are not aware of and then i kept asking 1. discomfort isn't just inevitable; it's essential. chase friction relentlessly.
2
1
1
516
Looking at the past 30 odd years of data you can see clearly the indices on an exponential rise. This means any reversion to mean nowadays will end up with bigger and bigger blips - the curves are getting steep. Just means we will see new records set soon…
1
6
843
They: If your project takes more than a week, it's already outdated. Soon, even two days will feel slow.. Me:
2
8
514
Today is the start of a new era of natively multimodal AI innovation. Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality. Llama 4 Scout • 17B-active-parameter model with 16 experts. • Industry-leading context window of 10M tokens. • Outperforms Gemma 3, Gemini 2.0 Flash-Lite and Mistral 3.1 across a broad range of widely accepted benchmarks. Llama 4 Maverick • 17B-active-parameter model with 128 experts. • Best-in-class image grounding with the ability to align user prompts with relevant visual concepts and anchor model responses to regions in the image. • Outperforms GPT-4o and Gemini 2.0 Flash across a broad range of widely accepted benchmarks. • Achieves comparable results to DeepSeek v3 on reasoning and coding — at half the active parameters. • Unparalleled performance-to-cost ratio with a chat version scoring ELO of 1417 on LMArena. These models are our best yet thanks to distillation from Llama 4 Behemoth, our most powerful model yet. Llama 4 Behemoth is still in training and is currently seeing results that outperform GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused benchmarks. We’re excited to share more details about it even while it’s still in flight. Read more about the first Llama 4 models, including training and benchmarks ➡️ go.fb.me/gmjohs Download Llama 4 ➡️ go.fb.me/bwwhe9
809
2,329
1,173
12,572
3,672,282
A billion-parameter voice AI is now open source Sesame AI just dropped CSM-1B - the voice engine behind its viral assistant Maya. - Fully open. - Apache 2.0 license. - No paywalls. No permission needed. Here's everything you need to know:
111
691
93
6,733
849,303