koplenko

@knveth This is the way. Watch the match while Sol Ultra burns through the weekly limit in the background 😂

There are people watching the world cup semifinal between the 2 best squads in the world instead of building businesses and empires with GPT 5.6 + Claude Mythos while trading assets on Robinhood L2 blockchain. Disgusting really

@francescoinweb3 The real advantage isn’t having the smartest model - it’s knowing which model to use for each task

Everyone is arguing about which model is the most powerful today. The company processes 10,000 customer requests every day. Instead of sending each task to the most expensive model, they have built a system of several AI. Llama parses simple requests, ChatGpt writes complex texts, and Grok checks data from the internet. It receives the goal, selects an executor for each task, checks the intermediate results, and decides what to do next: accept the response, send it for revision, or completely change the execution route. During the work, the team compared the performance. Before the introduction of this controller, about 35% of requests required manual processing. After implementing the single manager architecture, the number of errors decreased, the cost of processing a request was reduced to 0.05$, and the average response time increased. This is why the new role is the meta controller. Its task is not to write the best text or generate the best code. It must understand the entire picture and distribute the work among the team models, if the strategy fails, to rebuild the entire process. The most powerful model does not have to do the work on its own. It is much more valuable to use it where it has the greatest impact, as someone who controls the system and makes decisions.

n8n is about to lose its most important job. AI agents no longer need every step drawn as a node. They can choose tools, handle failures, and change the path themselves. The workflow is becoming a prompt. n8n will remain the plumbing - but it won’t be the brain.

@bridgemindai I think it really depends on the task. Fable still feels great for focused coding, but Sol gets interesting when tools and parallel agents are involved.

Claude Opus 5 is imminent. If Anthropic thinks this is the excuse to remove Fable 5, they are wrong. GPT 5.6 is not as good as Fable 5. But it is crushing Opus 4.8. Opus 5 has to be an INSANE jump from Opus 4.8 for me to keep my subscriptions. The bar is Fable 5. It has always been Fable 5.

@kimmonismus same here. The clearest sign 5.6 is different is how quickly I burn through the limits - I keep finding more work to delegate to it.

The wait is over. Release season has arrived. After a relatively quiet June, the pace has picked up again. GPT-5.6 was the breakthrough I had been hoping for. I’ve never burned through my rate limits so quickly. Fable 5 is back, Meta has released another competitive model with Spark 1.1, and SpaceX followed with Grok 4.5. According to SpaceX, however, their truly flagship model is still on the way. It looks like the Cursor deal is already paying off. As many know, the Fable 5 saga continues. Its availability in the subscription plan has already been extended twice, while Opus 5 has meanwhile become visible in Vertex. That strongly suggests Opus 5 will be released soon, likely as a replacement for Fable 5. Even more interesting, though, is that GLM-5.2 turned out to be the real “aha” and “wow” moment for many people. It made it clear that open source has now reached a level where, at least in key areas, it stands as a genuinely viable and competitive alternative to Western closed source models. What makes this even more surprising is that the company’s founder once again hinted today that their next model is also coming soon. Not long ago, during a discussion with Elon Musk, he also said they still plan to release their own Mythos-class open source models before the end of the year. As if that weren’t enough, leaks suggest that Kimi K3 will be released tomorrow. It’s worth remembering that Kimi K2.6 was, only a few weeks ago, arguably the most popular open source model around. Time moves unbelievably fast in the AI era. That was until GLM-5.2 captured everyone’s attention. Kimi K3 is expected to launch with a one million token context window, a significant leap that makes it even more compelling. In short, Opus 5, GLM-5.3 or perhaps GLM-6, and Kimi K3 all appear to be just around the corner. At the same time, the rumor mill suggests that ChatGPT 6 could arrive within a matter of weeks, featuring an entirely new pre-training pipeline. The long winter of waiting is over. The summer of acceleration has begun.

@MTSlive @cdngdev This is such a good example of why personal context matters. Turning a messy camera roll into something useful feels much bigger than just a wardrobe app

SITUATION EXPLAINED: Thijs gave GPT-5.6 Sol access to his camera roll and it built him a virtual wardrobe app on the spot. • @cdngdev gave the model access to his camera roll, it identified the exact pictures of every piece of clothing he owns from his own photos • He then had it generate new outfit combinations and render them on him using GPT image • @sama: "this would've been a whole startup not too long ago" @sodofi_: "I know three startups that do this and spent years building this, and they still haven't found product market fit, and I bet he did this with one prompt."

@DeryaTR_ @higgsfield_ai The model choosing the right image model is the underrated part. Once agents can route each subtask to the right specialist, the workflow matters more than any single model.

I recently started using @higgsfield_ai MCP with Codex for game graphic assets and I really like it a lot! I'll share some examples soon. The cool thing is GPT-5.6 Sol somehow knows how to choose the best model for graphics/sprites. So many cool tools and apps to use with Codex!

@jasonlk Scaling from 1 to 21 agents creates a new bottleneck: supervision. The next useful layer isn’t another agent — it’s an operating system for priorities, permissions, memory and exceptions.

Few things have been more fun in my career in software and tech than the past 12 months We've gone from 1 basic agent to 21+ AI Agents running all of SaaStr, much of SaaStr Fund, and so much more But the cognitive load is at the edge. It's at the limit. I don't think we can manage one more independent agent. And everything now is 24x7. The backlog. The agents never stop working. Mornings, nights, it's all agents, all the time. It's great, it's so cool, and I hate the term "burn out." That called life when you are building. But it's a level of constant cognitive load I've never really experienced before.

@dustinhollywood @OpenAIDevs @OpenAI @stages_ai Curious how Sol handled context drift once the project became large

Shout out to @OpenAIDevs @OpenAI Codex and all those resets on 5.6 Sol Ultra. I’ve been running it non-stop for 72 hrs 👀 wait till see what I put in @stages_ai I had a dream and shit got crazy.. 🔥🎥

@bitfalls This solves a real power-user problem. Once you have dozens of parallel threads, retrieval becomes part of the agent architecture — not just a sidebar feature.

The codex sidebar has been the bane of my existence as a power user. So here's a small tool to catalog and search them with additional metadata rather than spamming titles with PR numbers. Saves a ton of agent tokens on thread search. github.com/Swader/catalog…

@carlrichell how’s the fan noise once the GPU is actually under load?

My new mobile daily driver! It's a perfectly balanced laptop. Thin, light, powerful CPU and GPU, and battery life that gets me through the day. And that incredible 2K OLED display. system76.com/laptops/adder-…

@yacinelearning finally a kernel intro that doesn’t assume you already speak CUDA 😂

very neat beginner friendly intro to GPU kernel check it out!

@lukas_m_ziegler does it keep correcting the intercept point after the ball is released, or commit to one trajectory?

Catching flying objects with magnetic levitation! 🫴🏼 A team from ETH Zürich demonstrated catching a thrown ball using an electromagnetically levitated object. The levitator has full 6-degree-of-freedom control, can position and orient the ball in any direction. It predicts the ball's trajectory and intercepts it mid-flight. Two off-center magnets integrated into the levitator. That's what enables complete 6-DOF control instead of just hovering vertically. Traditional approach uses an analytical model (Multipole Expansion Model) that makes strict assumptions about symmetry and needs careful calibration. Their new approach is a bit different. Neural networks that learn the magnetic field relationships directly from data. Multi-layer perceptrons that capture nonlinear behavior while preserving the linear relationship between currents and forces. There are no symmetry assumptions, and it works with irregular geometries. Paper: arxiv.org/abs/2602.06618 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com

@digiii @trymelius the node canvas is the part i’d want to see — creative agents get messy fast once there are five steps instead of one

Be @trymelius (Melius): The creative harness for agents. An AI-native creative harness for agents - A platform where you direct AI agents to build generative workflows (images, video, campaigns, etc.). - Slack agent integration for fast ideation-to-output - Node-based canvas for multi-step creative processes - Access to top models (Seedance, Kling, Veo, etc.) - Strong focus on character consistency, multi-shot video, and professional workflows Turns you into the “creative director” while agents handle execution - aimed at marketers, agencies, and creators. Founders: Joowon Kim (CEO & Co-founder) - @n0w00j Arnav Ramu (COO) - Young Kim (CTO) - @seedelano

@antpalkin the speed matters more than the AGI date tbh — most people are still planning around normal software cycles

Demis Hassabis, Nobel Prize winner and CEO of Google DeepMind, just called it: AGI is a few years away. His exact words yesterday: "we are standing in the foothills of the singularity" Everyone's sharing the safety parts. Almost everyone skipped the one line that matters for your money: "10x the Industrial Revolution at 10x the speed" Run that math. The last Industrial Revolution minted Rockefeller - $400 billion in today's money, the richest man who ever lived. It took him 40 years of railroads and oil fields. Hassabis is telling you the next one compresses that into a few years, and it's made of software. And new money always shows up in markets first. No factories to build, no permits, no supply chains - just data in, positions out. The "recursively self-improving systems" he says the world isn't ready for? Early versions already trade against you every day. That gap between people who run them and people who don't is exactly what he means by 10x speed. He wrote this essay to ask for guardrails before the wave hits. Read it twice and you'll notice what it really is: the most credible man in AI telling you the wave is already visible from shore. Bookmark this. In a few years, "I read Hassabis in the foothills" will sound like "I bought Bitcoin in 2013" - and you'll want proof you were here. I write about this frontier every week - AI agents and what they're doing to markets. Follow me, stay in the AGI foothills.

@Aspect0F 100%. It saves me a lot of time

@koplenkoo that honestly sounds like a complete game changer for your workflow.

I asked GPT-5.6 Sol to review a Voice AI-agent I’m building. Ultra split the job: one agent checked the prompt, another traced the call flow, and another searched for failure cases. Normally, I do those reviews one by one. Sol returned everything as one result. That was the first time AI felt like a team, not a tool.

@0x_fokki the 60 hours inside a 24-hour day really puts it into perspective.

@koplenkoo Makes sense

OpenAI’s heaviest 1% of internal Codex users now run more than 60 hours of agent work per day. Not per week. Per day. They do it by running multiple agents in parallel. Pause at 0:07 — an entire team of specialized agents from a single workspace. And this has already moved beyond engineering: Legal, Finance and Recruiting now use Codex as their primary AI tool too. This is the shift most people still miss. The advantage isn’t getting one better answer from a chatbot. It’s turning one goal into multiple delegated tasks, letting agents work in parallel, then reviewing the results. AI literacy used to mean knowing what to ask. Now it means knowing what to delegate.

@tolik12308 exactly — the unit of work is no longer limited by your own time.

@koplenkoo This is the shift that changes the rules of the game

@kocer_eth yeah, at that point execution stops being the bottleneck. reviewing everything does.

@koplenkoo running multiple agents in parallel for delegation is absolutely crazy for scaling output