🚨New Must-Learn Magik Combos! DP Loops + fireball loops! These are high damaging combos that also set up some of her best oki. This character is way too fun lol Here are a few examples. Complete Magik Guide Soon! #Tokon_Magik# #MARVELTōkon#

It's so magical when you wake up to a difficult PRD being completed correctly, thoroughly, and ready for landing. Man, it's just magic. What a time to be alive. Still can't believe this all comes from inferring the next token.

if you need to get your infra spouse’s attention here are the magic words for the summer: - time to first token - net flops utilization - platform fungibility - spike capacity next week use with caution

Dominic Kundel (from @OpenAI) gives the inside scoop on where GPT-5.6-Sol gets magical: computer use. Background browser tabs, app control, multi-agent fanout, and Codex verifying its own work all change when latency drops. Join us in the Token Billionaires Lounge, presented by @cerebras and @aiDotEngineer. In conversation with @dkundel // @MilksandMatcha

so far i think our claude code plugin sucked. now it is magical it was not very clear when it was being helpful it felt like it's kinda ruining the context It's highly loved and recepted, we have thousands of people using the claude code plugin every day, with many stars on Github. Really wanted to make it good. I sat on it for an entire day and tried to figure out every single way to deliver the value with our plugins. and I think we finally have an answer now: 1. Showing the value, somehow... Because we can't control the UI in claude code, we figured out many different ways to show when supermemory is being useful in the session. Now, supermemory will show it's value, in: - User prompt submit - Statusline - Inline, with the ◪ icon. So when you see ◪, you know supermemory is doing it's magic. It also shows the number of tokens injected, so you can be confident about pollution. 2. Great memories Coding memories are different, and honestly, the learnings that the older version was doing were not being helpful. It would remember very arbitary stuff like "The user used git push" (yeah that's terrible. 🤮). We have updated both the engine and more harness level changes to make sure that the learnings are great for coding as well. Now, we try to infer the source of truth for some things, so if something is obvious in the code (this is a typescript project), there doesn't need to be any memories for it. The memories should be your behaviours, preferences and other things. We do this through profile buckets and entityContext in supermemory. 3. Experience and observability Ok. So the memories are good. But how do you bring it back into the harness? We have both hooks and tools, and the tools are now served via a local MCP, that the agent uses pretty decently often now. This is to gather additional context. how do we show it to people? Now, the claude code plugin (and other plugin) transcripts and learnings show up in the supermemory app in the agents tab. Here, you can see the full graph of what is learnt, inferred, and how supermemory is learning over time. in a beautiful way, so you can observe and make sure that supermemory is doing it's job right! 4. Failures and when things go wrong. This is one more thing that I really really care about. While testing the plugin, I happened to be in an airplane with really slow internet. Seeing that supermemory was making the agent feel slow as well (since it's a network call before it starts answering), the plugin now guarantee takes less than 500ms (or else it is skipped) and all errors are failed open. This means that even if supermemory fails your work will not be affected. One more thing that was concerning me was honestly the token count. How many tokens is supermemory injecting? is it polluting the context? Now we clearly show the token count so you can be sure it's doing good. More often than not, it will always be less than 100 tokens. 5. Feeling the difference As I was showing this to my friends after feeling the magic myself, a common question was "How is it different from claude code's default memory?". Yes, the memories work cross-harness. That's one plus. But that's not it. If you open the "Auto-memories" folder in claude, you'll notice that it barely learns anything. Things are only learnt if you are visibly angry or direct about "Don't do this". Then, it creates a file with "When doing this, do XYZ" or "Avoid XYZ" There's barely 10 files in my repo, which I've been working on for more than 2 years now and using claude code since last 8 months. There are also a lot of outdated ones that are no longer even necessary. In order to use these memories, claude has to load the whole files automatically, which it never does... Supermemory learns and captures even the small implicit details about how you would want the agent to behave, and they are always fresh, since forgetfulness is built in. It is also used and picked up by the agent, automatically. you should REALLY try our new claude code plugin. Really proud of the entire team's work in making this amazing. these changes are coming to other plugins soon

#ILoveTokyo# And I know so many of you do, too. Thirty years ago, as our family prepared to leave New York City for Tokyo, my daughter was given a beautiful picture book about the city she was reluctantly leaving behind. It was a comforting token, and immediately, I wished I had a similar book about Tokyo to show her the magic, wonder, and vibrant adventure that awaited her on the other side of the world. There was none. And to this day, there is none. For years, I pitched the concept to major publishing companies. Time and time again, the answer was the same: "The illustrations are too detailed." "It's too expensive to produce for a children's book." They wanted simpler drawings. They wanted shortcuts. But I refused to compromise on creating a book that children would refuse to put down—a keepsake they would cherish and revisit for years. I tried crowdfunding for the project, but I fell short of the target. Yet, as the saying goes, when one door closes, another one opens. Thanks to a wonderful introduction, I connected with Ryoko Yomiuri Publishing Company—a team that truly understood my vision and appreciated the rich, detailed artistry I refused to give up on. Together, we are finally bringing this 30-year dream into children's hands! #ILoveTokyo# will officially go on sale on November 18th!

So many things happening on solana right now, it's hard to keep up, but Sensei got you - @phantom rolled out football world cup prediction markets in-app on Solana - @Backpack announces no fees on stock trading through July and private beta has gone live for the backpack credit card - @FlashTrade V2 is now out of beta with fees as low as 2 bps and one click trading. Trade with link in their bio. - @blknoiz06 saving solana trenches with over $9M distributed in airdrops to 700+ wallets so far on his coin ANSEM - @MagicEden is facing a federal class action that accuses ME of misleading buyers of ME, a token down 99% from its 2024 launch. - @Titan_Exchange announced new DCA badge that can be obtained through trading $1k in DCA volume - @solsticefi postpones 2nd SLX unlock from july 5th to july 9th with $2,000 in SLX giveaways every day as incentives for delay.

MiMo-V2-Flash is live. It’s just step 2 on our AGI roadmap, but I wanted to dump some notes on the engineering choices that actually moved the needle. Architecture: We settled on a Hybrid SWA. It’s simple, elegant, and in our internal benchmarks, it outperformed other Linear Attention variants on long context reasoning. Plus, a fixed KV cache just plays way nicer with current infra. Note: Window size 128 turned out to be the magic number (512 actually degraded performance). Also, sink values are non-negotiable—don't skip them. MTP (Multi-Token Prediction): This is underrated for efficient RL. Aside from the first layer, it needs surprisingly little fine-tuning to hit high accept length. With a 3-layer MTP, we're seeing >3 accept length and ~2.5x speedup in coding tasks. It effectively solves the GPU idle time from long-tail samples in small-batch On-Policy RL. We didn't get to squeeze it into the RL loop this time due to deadlines, but it’s a perfect fit. We open-sourced the 3-layer MTPs so you can develop with it. Posttrain with MOPD: We adopted On-Policy-Distillation from Thinking Machine to merge multiple RL models, and the efficiency gains were wild. We matched the teacher model's performance using less than 1/50th the compute of a standard SFT+RL pipeline. There’s a clear path here for a self-reinforcing loop where the student evolves into a stronger teacher. Huge props to my team. They sculpted these ideas from scratch into production in just a few months. Full breakdown is in the tech report. If this kind of pragmatic engineering resonates with you, we should talk.

New Stanford paper argues that, under equal reasoning budgets, one LLM usually solves multi-hop problems better than many coordinated ones. The core point is almost embarrassingly simple. A single agent keeps the whole problem in one internal chain of thought, while a multi-agent system has to slice that chain into messages, summaries, and handoffs. Every handoff is a compression step. And once reasoning is compressed, some information is easier to drop than to recover, which is why the paper leans on the Data Processing Inequality as a formal explanation rather than just an empirical hunch. The experiments back that up across Qwen, DeepSeek, and Gemini on FRAMES and MuSiQue: when thinking-token budgets are matched, single-agent systems usually match or beat sequential, debate, role-based, and ensemble setups. Here’s the part most people miss. Many celebrated multi-agent gains may not be architectural gains at all. They often come from spending more test-time compute, surfacing more visible reasoning, or benefiting from evaluation quirks that make the pipeline look smarter than it is. The paper is especially sharp when it looks for the boundary case instead of pretending the rule is universal. When the single agent’s effective context is degraded by masking, substitution, or misleading distractors, multi-agent pipelines become more competitive and sometimes win, not because message passing is magical, but because structure can partially stabilize corrupted reasoning. That is a much narrower and more useful claim than “more agents is better.” It suggests the real trade-off is not single versus multi so much as latent reasoning versus external coordination, with context quality and compute accounting deciding which side looks stronger. For multi-hop reasoning, the default should now be clear: start with one strong model, and treat extra agents as a repair strategy, not an upgrade. ---- Paper Link – arxiv. org/abs/2604.02460 Paper Title: "Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets"