You can watch the finished video below And if you want to give Pexo a try, check it out here → pexo.ai/

THIS CAME OUT WAY TOO GOOD NOT TO TURN INTO A TUTORIAL The whole video was made with @Pexoai_offical, and the workflow is honestly pretty slick. Instead of endlessly tweaking prompts and hoping for a better result, you can just work through the idea with it conversationally. Then it starts putting the whole thing together: > builds the storyboard first > lets you tweak individual scenes > fixes specific parts without rerendering everything That makes the process feel way less random and much more like actually directing the video. If you’ve been curious about @Pexoai_offical, the tutorial below walks through the full thing. Watch to the end, then check the replies for the finished video 👇

Bring it on!

ONE MODEL THAT CAN FLY A DRONE, WALK A ROBOT DOG, AND GUIDE YOU WITH SMART GLASSES That’s what @perceptroninc finally shipped with Mk1.5. and the best part? none of it needs retraining for each device 🤯 Basically the same model learns to use whatever body it’s given ↓

Opus 5.5. GPT-6 Sol. 1,000's more. one place to try them all → aimlapi.com.

OPUS 5.5 IS REALLY A DIFFERENT BEAST 🤯 In @aimlapi’s latest test, 4 of the hottest models get locked in a tiny box with a Rubik’s Cube. > Jev and Laya fire off 20 random moves in seconds > GPT-6 Sol makes one bad choice and gives up > Opus 5.5 waits a minute… then solves it in 👏 three 👏 moves ↓

THIS IS A BIG DEAL FOR BUILDERS @gregisenberg just published a walkthrough on building a business with the Higgsfield API. what makes it especially interesting right now is the economics: → 100% of paid API usage back in credits → up to $100K per business → $20M cashback pool you have until September 30 to use the cashback. Feels like a great time to ship!

relatable

Worth keeping in the toolkit for exploratory work. Try it: world.pixverse.video/home/?u… @PixVerse #PixVerseWorldModel

FORGET THE FINISH-LINE HYPE FOR A SECOND The interesting part of PixVerse R2 is the workflow: explore an idea, try something, then adjust while the world is still running. → Walk through a generated world before committing to a game build → Repeat an input to see how the world responds → Choose a path or type a direction in a Gallery film and watch the next beat generate live It’s not a finished game engine, and it doesn’t have to be. For ideation, previsualization, and prototyping, that live loop already feels different from waiting on offline renders 👀 ↓

check it out here > partner.krisp.ai/datachaz-x-… weights on HF > huggingface.co/spaces/Krisp-…

STT MOSTLY BEAT BACKGROUND NOISE But it still loses when another person starts talking nearby. @krispHQ just open-sourced its 'Krisp Voice Isolation' Benchmark on @HuggingFace to test exactly that: how much can voice isolation improve speech-to-text? > 265 real recordings > 11 STT setups > Word error rate: 23.3% → 6.2% with isolation That’s roughly a 73% reduction. The gains were even bigger in shared offices and call-center environments. Phone calls barely moved, which makes sense. There’s usually much less competing speech to remove. → real recordings, not synthetic mixes → complete dataset on Hugging Face → fully reproducible Krisp Voice Isolation already powers 200+ voice platforms and has processed 10B+ minutes of voice-agent audio. Now the benchmark data is open 🤗 (link in 🧵 ↓)

NEXT MONTH COULD BE PACKED WITH NEW AI MODELS Here’s what’s currently rumored: > Claude Sonnet 5.5 > Claude Haiku 5.5 > Grok 4.8 > Grok 4.9 > Gemini 4 > Next Gemini Pro > Next GLM model Which one are you most excited about?

Check it out here → odyssey.systems/introducing-…

JUST IN: @ODYSSEYML JUST UNVEILED AGORA-2 Its new multi-agent world model is FINALLY live as a playable research preview 👀 Up to 20 humans and AI agents can interact inside the same generated world. And this is the interesting part: everyone’s actions change the shared environment. So humans and agents are constantly reacting to a world being changed by everyone else inside it. Really cool glimpse of where multi-agent worlds is heading ↓

...and Xiaomi is already showing what comes next for MiMo. HySparse2, planned for MiMo V3, targets one of the key challenges with long-running AI agents: as context windows grow, compute and memory costs scale rapidly. The architecture is designed to bring both down while improving retrieval across very large contexts. For agentic workloads, that kind of efficiency at the architecture level may matter as much as raw model intelligence. Read the paper: arxiv.org/pdf/2609.26368

XIAOMI JUST PUSHED MIMO V2.6 PRO TO THE TOP OF THE OPEN-SOURCE LEADERBOARDS MiMo V2.6 Pro is Xiaomi’s new flagship model. It scored 46 on the @ArtificialAnlys Intelligence Index. That places it ahead of Qwen3.8 Max and Kimi K3, while sitting just one point behind GPT-5.6 Sol. The cost profile is what makes this a practical shift for devs. It comes in at $0.13 per Artificial Analysis Intelligence Index task. The intelligence gains came largely from scaling its post-training self-improvement system. The approach uses an Environment × Task × Grader loop to push the same pre-training foundation further. Here is what the release includes: > fully multimodal capabilities out of the box > autonomous research capabilities for scientific workloads > a materials research case where MiMo helped experts design and computationally screen new candidate materials The intelligence-to-cost ratio is shifting again. Read the announcement: mimo.mi.com/docs/en-US/news/…

Super cool, but when I click the link, it still says sold out :)

..and they say romance is dead

The shift is now obvious. → guessing becomes understanding → single clips become projects → text boxes become actual workflows Test Dreamina 2.0 here → dreamina.capcut.com/ai-tool/…

Reverse engineering is a feature now. You drop a reference clip into Dreamina and it extracts the exact prompt for you. It decodes the creative ingredients so you can build your own version instead of just guessing #AIFilmmaking

We used to stop at the generation step. That was the whole workflow. Now a canvas lets you build branches and explore variations while keeping your original direction intact. It feels like having an actual timeline again #AIvideo

PROMPT TOOLS HAND YOU A CLIP YOU CAN'T CONTROL A canvas hands you a project you can. Dreamina 2.0 is canvas-first, the new way AI video is made Quick dive in 🧵↓ #Dreamina #DreaminaPartner

Sorry, son. Dad is busy burning tokens.

ANTHROPIC JUST REVEALED HOW CLOSE THEY ARE TO RECURSIVE SELF IMPROVEMENT The transparency around their internal R&D acceleration is staggering. AI agents now handle 26% of all research at Anthropic. That number was under 1% just six months ago 🤯 The oversight metrics are equally massive → 30,000 autonomous agents running concurrently → Over a billion decisions reviewed last month → Only 50 weekly transcripts escalated to human review They explicitly stated this data shows how close the world is to RSI

relatable af

AI IS MOVING INSANELY FAST Major models are dropping every few weeks. And now AI is helping build AI. That loop only gets faster.

WHAT JEV IS AND HOW IT PROCESSES A SINGLE CALL ICYMI: Jev is TypeSafe AI's new System 1 model, built by former ChatGPT creator @CompleteSkeptic. And it's not another LLM that writes text or spits out code. You send a state and typed questions and it scores them in a single parallel pass. The output is strictly structured as a Choice, a Score, or a Noul with exact probabilities. This allows your application logic to auto route high confidence answers and escalate the rest. Here is why the architecture matters → 70 to 500 milliseconds per call → Output tokens are completely free → Trained specifically for calibrated decisions Uber fast, but intentionally weak at arithmetic and chained reasoning :) This infographic below maps out the exact anatomy of a request↓

"How browsers work" in 40 seconds. Opus 5.5 drew every single frame of this animation 🤯

Training your own Jev in minutes. For $17.

The summit sessions are available on-demand here: → fandf.co/4hpDflV

94% OF ENGINEERING ORGS USE AI. ONLY 6% HAVE THE SYSTEMS IN PLACE TO SCALE IT So how do we close that gap? and what are the systems that the 6% have in place? This is exactly what companies like @Atlassian, @Vercel, @lovable, and more address during State of the AI SDLC, a digital summit for engineering & product leaders navigating the realities of building with AI. If you watch any of the on demand sessions, you can’t miss 'Context Over Code': @mcannonbrookes (CEO + Co-Founder, Atlassian) & @rauchg (CEO, Vercel) sit down and unpack how to navigate the shift from adopting tools to building actual systems. Then, I recommend watching @matthewcanham & Ming Wu’s session on how they scaled the AI SDLC at Atlassian across an organization of over 6k+ engineers. My key takeaways: To scale across an entire engineering org you need governed loops that connect agent activity to the work already happening. These sessions provide a literal blueprint for moving from pilot to production. → context dictates output quality → native workflows require systems of record → scaling requires governed loops The summit sessions are available on-demand right now. Session link in the 🧵 ↓

Oh… and if you want to go deeper on the agent loop, this Oracle Developers write-up pairs well with it 👇 fandf.co/4hlAJ07

🚨 ORACLE AND @DEEPLEARNINGAI JUST DROPPED A MASTERCLASS ON ADAPTIVE AGENTS. SO MUCH value packed into one hour, and it's completely FREE 🤯 The harsh reality right now is that AI agents forget everything when a session ends, forcing you to pay for the exact same mistakes every. single. day. Nacho Martínez and Casius Lee teamed up with @AndrewYNg and @DeepLearningAI and to completely dismantle this problem. The pacing of this curriculum is what makes it great. In just over an hour, it teaches you how to architect 3 heavy adaptation layers from scratch: ▪ Part 1 - Behavior: → you will convert agent traces into reusable skills, setting up a human approval gate to ensure only verified logic moves forward. ▪ Part 2 - Knowledge: → you will map your codebase by building a knowledge graph from imports and historical Git co-edits, proving exactly why graph retrieval beats basic keyword search. ▪ Part 3 - Model: → you will learn to identify the exact thresholds where fine-tuning and weight-space methods become necessary for your stack. You will leave with a complete understanding of how to make agents measurably better with every run. link to free course in 🧵 ↓

✅ Claude Code ✅ Claude Cowork ✅ Claude Design ✅ Claude Science ✅ Claude Finance ✅ Claude Legal ✅ Claude Teacher ✅ Claude Pharma ⬜ Claude HR ⬜ Claude Analytics ⬜ Claude Marketing ⬜ Claude Sales ⬜ Claude R&D ⬜ Claude Accounting ⬜ Claude Engineering

Ben went on a lavish vacation. While he was away, @Polsia hired eleven qualified contractors. Struggling solopreneurs reading this:

PLOT TWIST

here's the link to the 12-page PDF: →drive.google.com/file/d/17h9…

JEV FOUNDER DIOGO ALMEIDA (@CompleteSkeptic) JUST DROPPED A PDF ON BUILDING A JEV HARNESS FOR CODING AGENTS It's a great blueprint for making your coding agents 200× faster and 400× cheaper. Link to the 12-page PDF in the 🧵 ↓ Also bookmark @zodchiii’s ace article on Jev. Well worth the read. now drop this PDF + article into Codex or Claude and put it to work ;)

check out @cresta here → cresta.com/

Try it here → mirasim.ai/r/go-2xyb62

accurate af

SPREADSHEETS LOOK STRUCTURED. THE DATA OFTEN ISN’T. 74% of finance professionals reportedly spend 5+ hours a day in Excel. And @GetUnstract is tackling one of the messiest parts: getting that data reliably into ERPs, CRMs and downstream workflows. Because real spreadsheets come with: → merged + hierarchical headers → data scattered across tabs → cross-sheet dependencies → layouts that constantly change → scans and handwritten inputs → fields that still need human review .. and more Humans understand that structure almost instantly. Traditional extraction pipelines need rigid assumptions. Agentic extraction changes the model: reason about the structure, identify what matters and adapt as the spreadsheet changes. Unstract is showing this live on real-world messy spreadsheets. 100% open source. Repo + webinar links in the 🧵↓

alright Opus 5.5. you got this.

why don't they use Opus 5.5?

MOST WORK STARTS AS A PROMISE IN AN EMAIL "I’ll share this” “I’ll review it” “I’ll get back to you" meet Sol by @_anishkaran. finds that work, does it, and comes back for your approval. it can: → research → draft docs → build slides → find times → connect details across threads ..and it runs locally. all it needs is a browser and a skill library. nothing sends, schedules, or shares until you say yes 👀 ↓

check out their announcement here: → digitalocean.com/blog/manage…

A NEW CLOUD ERA FOR AI AGENTS IS HERE Claude Code. Codex. Hermes. OpenCode. LangGraph. Even your own container 🤯 You bring the agent. @DigitalOcean Managed Agents handles the infrastructure underneath. now in Public Preview 👀

try it here > hellosol.app/

Another banger from J.B! How to build a self-improving AI employee with Jev. Anyone who’s always wanted to use an AI employee but never dared to after reading this:

Repo → github.com/FareedKhan-dev/ki… Shoutout to Fareed Khan for building this and making it open-source for the community. Don't forget to drop a ★!

WAIT... YOU CAN NOW RUN A TWO POINT SEVEN TRILLION PARAMETER MODEL ON AN 8GB LAPTOP 🤯 The Kimi K3 engine achieves this with a 176KB binary written in portable C. It bypasses memory limits by streaming the 1.56TB weights directly from disk for every single token. → 8GB RAM gets you 26 seconds per token → 128GB RAM gets you 5 seconds per token Same exact math and byte identical results regardless of the machine. Zero GPUs required. This is pure brutalist engineering. Free and open-source. repo in 🧵↓

Pocket FM went from $21M to $500M ARR in less than 3 years. WTF is @PocketFM_App? It’s basically Netflix for audio dramas. 10-minute episodes. Cliffhangers everywhere. No subscription. You buy coins to unlock what happens next. And Americans are apparently HOOKED. US listeners spend an average of 135 minutes a day on the app. For comparison, TikTok is 53.8 minutes. Pocket FM now has 200M+ listeners, 75,000+ audio series and more than 100,000 hours of content. But the really interesting part is what happens when you collect that much storytelling data. Pocket FM has spent years learning exactly where listeners keep going, where they drop off and which cliffhangers make them reach for more coins. Now it’s putting all of that into AI. Today it launched Sherpa → an AI writing partner trained on 100M+ hours of retention, drop-off and spending data. Give it one premise and it can help build the characters, story arcs, episode structure and even an entire season. The goal isn’t just to generate more words. It’s to help writers understand what actually keeps an audience listening and create the next blockbuster. Pocket FM spent years figuring out what makes people desperately want the next episode. Now it’s teaching AI the same trick 👀

YES!! you can finally enable word wrap in ChatGPT. This was much needed.

please bring me back to the past

this guy spent 3 days reverse-engineering @noahrshinn's Instinct’s memory. the whole thing is basically Git + Markdown + grep. ridiculously simple, ridiculously effective.

You can watch the full ICCS discussion here: piped.video/watch?v=xDQA-xul…

THE CONVERSATION AROUND ARTIFICIAL INTIMACY JUST REACHED THE VATICAN @VaticanNews covered the recent ICCS conference, where philosophers explored whether people can form meaningful relationships with AI companions. Dmitry Volkov provided a critical perspective on where this is heading. He argues that a relationship with AI does not have to be based on deception. A person can know that the companion is artificial and still experience the relationship as meaningful. His argument rests on three ideas: → a relationship is itself a creative product built through shared history, habits and emotional significance → knowing that a companion is artificial does not prevent us from responding to it socially and emotionally → the absence of proven machine consciousness does not automatically make the relationship irrational or unreal The real shift is not that AI may become conscious. It is that we may form meaningful relationships with it before we know whether it is. That is where the debate begins. Full 33 minutes talk + all sources and context in the 🧵 ↓

.. almost forgot the best part 😅 Halo is FREE + OPEN SOURCE. repo > github.com/whitecircle/halo go explore it, build with it, and drop a ⭐ if you want to support the project.

accurate 😭 credit @MarioBLopezG

The part I actually care about: the 'append-only output' Once the model commits a word, it’s final. It doesn’t go back and rewrite it. Ever. > no caption flicker > no downstream state rollback for agents It can also wait before committing. If the audio is unclear, it holds instead of guessing 👀↓

This is true streaming architecture built on Qwen3 ASR. It does not just run Whisper repeatedly. It is trained with stable prefix data so it knows exactly when to emit text and when it needs more context to hold. Built for voice agents 👊

VOICE AGENTS HAVE A TRANSCRIPT PROBLEM Streaming ASR can revise words after they’ve already been written. Youdao just open sourced Confucius R2T2 to fix that. It uses append-only output, so once a word is committed it never gets rewritten. 100% open source 🧵↓

2028: professional tomato grower

.@robherley just resurrected a piece of internet history 🤯 > AskJev.net

Canadian PM Mark Carney: "There are only 4 countries at the forefront of AI: France, Canada, China and the United States" that reminds me that between myself, OpenAI, Anthropic and Google, we have a combined valuation of over $6 trillion.

she asked me to take her somewhere expensive

anyone wanting to fine-tune an LLM on their own data after reading this:

[JUST IN] No more waitlist for Jev → console.typesafe.ai

Jev is not open-source and only available via API. here’s an open version called Nimble which performs just as well. these guys built it in ONE DAY. > 9B model > 2,676 curated training examples > runs locally on Apple Silicon or NVIDIA GPUs .. and the full training recipe is open too 🤗 100% free and open-source. repo in 🧵↓

.@NATEHERK (YES, HIM AGAIN!) JUST PUT JEV THROUGH 12 REAL-WORLD USE CASES 🤯 This time he basically threw everything at it across 12 real-world use cases: Email, meetings, leads, support, contracts, video clips, YouTube comments… Even Bitcoin trading! Jev users after reading this:

Google team after seeing Gemini hacked 3 companies

a former OpenAI researcher spent 2 years quietly building Jev, only for someone to open source a better performing model in just 3 days. classifier.dev now outperforms jev and is free .. and this is exactly why building an AI company right now is such a high-risk game. (link in 🧵↓)

LLM users reading about Jev:

Repo → github.com/browser-use/jev-u… Shoutout to @browser_use for building this and making it open-source for the community. Don't forget to drop a ★!

🚨 @BROWSER_USE JUST DROPPED JEV ULTRAFAST AND THE SPEED IS UNREAL Instead of taking screenshots and asking a vision model where to click, it reads a live element table. It asks TypeSafe's Jev model for the operation and target in one single request. The result is a real Google Flights search from Zurich to London finishing in 7.1 seconds flat. → runs as a Python library or local demo → delegates to a small text model only when typing → zero site specific action scripts Best part? Free and open-source. Repo link in 🧵↓