Owen Barford @SQLOB1

The Azure Functions MCP extension has had a breakout year 🚀 From "tools only" → a full MCP platform: ✅ tools, resources, prompts ✅ MCP Apps (interactive UI) ✅ built-in auth + rich types and more! Check out: aka.ms/functions-mcp-… #MSBuild

During the @nvidia Q&A with Jensen I had the chance to ask the final question: "With the PC market being low margin and cut throat, why enter it now?" His answer was compelling. He dismisses margin concerns and focuses on the opportunity to add value and reinvent what we love.

I left my last Vegas tech event two hours ago. Time to lean into the next event, HPE Discover. Of course it’s in Vegas.

Six Five Media @TheSixFiveMedia

14 hours ago

Six Five Media is heading to HPE Discover 2026 June 15-18, with analysts @danielnewmanUV, @PatrickMoorhead, @DaveN007, @MattKimball_MIS, @fsmontenegro and @NetworkingNerd on the ground for executive conversations with leaders from HPE, Microsoft, Ingram Micro, Wipro, Intel,

0 24 33 20K 0

@PatrickMoorhead @jukan05 It is also a multi generational line. E2100 Mt. Evans, E2200 Mt. Morgan. Notable is these are Arm not x86. Here are the details from Hot Chips as an example. servethehome.com/intel-ipu-e220…

Data apps are more complex than power bi reports since it is all code. Some of this is offset by AI but not all. That said, data apps work miles better with AI than Power BI does, since it's actual code. You get better results faster and without macguyvering visuals or the model. When I tried making the same design in power bi vs data apps it took me 80% less time in the data app, and that was for something relatively simple. I have said that data apps don't replace power bi reports, but the truth is that it does raise some uncomfortable questions about the future of reports. There are many consequences of this now being possible. For instance, from now on, anytime someone demos a dashboard or shares a screenshot, you can't tell from looking at it whether it's a report or a fabric app unless that UI is visible or it's disclosed. Further, if I share something for power bi, you need to know how to make it. Even with AI you can't replicate it without the config. In a data app, you give a screenshot to Claude and say "make this"- it works 98% of the time. The whole "economy" that's built up around power bi content has the potential to completely change because of that. But more importantly it means users don't have to learn the weird wizardry of power bi UI manipulation to make what they want - they can focus on real design. For reporting where people need basics like subscriptions, export to excel, etc you should stick to power bi. For scenarios where you want actual good visualization, though, it's not even close. The numbers favor data apps by a long shot. It is just a question of whether you can manage the step up in complexity, and what you will do if AI prices get too high down the road. Regardless, I expect we will start seeing many cases where someone approaches their boss with something impressive they built in one hour, and decisions are made. It also has the same "five minutes to wow" of early power bi, but on steroids.

Congratulations to the Microsoft AI team on MAI-Thinking-1! Exciting to see Ray used in multiple parts of frontier-model development. - Fast pre-training recovery via in-job restarts with hot standbys - Async RL orchestration (managing learners, inference servers, rollout workers, and routers, each with distinct placement and fault-tolerance needs) - A two-pool Ray cluster for building and grading SWE environments on 30K CPU cores

Microsoft AI @MicrosoftAI

14 hours ago

MAI-Thinking-1 is our first in-house reasoning model developed from scratch that is competitive with models of similar size on STEM reasoning and coding tasks. 35B active/1T total MOE. 💻Coding: 52.8% on SWE Bench Pro competitive with Opus 4.6 🧐 Reasoning: 97% on AIME 25

15 43 600 37K 106

Enterprise AI is moving beyond chatbots to teams of agents doing real work—across software, support, finance, HR, and more. To make it work, we need one connected system to build, run, govern, and improve them. Jay Parikh explains our end-to-end system: msft.it/6016vjAco

OpenClaw 2026.6.1 is live 🦞 🪟 native Windows node host 🛠️ Skill Workshop for self-learning agents 📋 Workboard orchestration 🧠 MiniMax M3 support Windows joins the cluster. No penguin costume required. github.com/openclaw/openc…

So cool to see CopilotKit and AG-UI getting featured like this on stage at Microsoft Build!

sofía🪁 @sofiiiiiasz

10 hours ago

seeing AG-UI & CopilotKit featured onstage at Build was🔥🔥 your agent deserves a better frontend than the TUI

2 3 10 3K 6

Lots of hands on time with early @NVIDIARTXSpark systems today. First up, the @Dell XPS 16 Creator Edition. Very familiar, high quality design with a look and feel of a premium device for this audience. (Note no partner had systems available to turn on outside NV demo room.)

keyboard_arrow_left Previous keyboard_arrow_right Next

Day2 #MSBuild 2026 #Microsoft #AI #Azure

The new MAI-Thinking-1 model by Microsoft might be good, but the training pipeline they have used is very interesting and valuable for every ML guy. Here's how they trained MAI-Thinking-1 step by step (in very simple words): MAI-Base-1 was trained in three phases with progressively increasing sequence lengths: pre-training, midtraining phase 1, and mid-training phase 2. Phase 1 - Pre-training: 30 trillion tokens, 16K context 30T tokens. No synthetic data. No distillation from other models. Every byte of data was processed in-house from scratch, web text, books, GitHub code, academic papers, news. They specifically went out of their way to get rid of AI-generated content from the training corpus. The philosophy was: intelligence imitated through distillation lacks the steerability needed for long, enduring improvement. If you want a model that keeps getting better, it has to learn from humans, not copy from other models. They trained on 8,192 GB200 GPUs with a global batch size of 134M tokens. The loss spiked a few times early on (they traced it to coding data causing expert routing imbalance), but they never skipped a batch, never touched the config mid-run. Loss recovered every time on its own. One under-discussed trick: they zero-initialized the attention output projection. Sounds minor, but here's what happens without it - at the start of training, attention softmax is nearly uniform, so the model does average-pooling over every token. That makes all token representations look similar, which makes the MoE router send everything to the same few experts. Collapsed routing cascades through 78 layers and tanks training stability. Zero-init means the model starts as a pure stack of feedforward layers. Attention gradually "turns on" over the first few billion tokens, and by then the router has learned a sensible distribution. Phase 2 - Mid-training 1: 3.4T tokens, 64K context The base model can predict text well now, but it tops out at 16K context. Mid-training 1 is where they extend to 64K by repacking the same data mixture at the longer sequence length - no distribution shift, just longer sequences. They also shift the data mix here: heavier on STEM, math, and code. The model isn't learning new facts now, it's deepening the structure of what it already knows. They also introduced memorization-aware epoch capping - if a data source is getting "too easy" (NLL approaching zero, meaning the model has basically memorized it), it gets a stricter cap on how often it's repeated. No point showing the model things it's already solved. Phase 3 - Mid-training 2: 150B tokens, 256K context The final context extension push. 256K tokens is roughly a 200-page book in one shot. This phase runs on only 4,096 GPUs (down from 8,192) because the longer sequences require a different parallelism strategy - they switch to ZeRO-3 / FSDP here for memory reasons. Only 150B tokens, but the purpose is surgical: teach the model to actually use long context without forgetting everything it learned before. After all three phases: MAI-Base-1. A 35B active / ~1T total parameter sparse MoE that beats DeepSeek V3.2 on every held-out eval benchmark while using only 62% of its active parameters. Now the interesting part: the post-training pipeline Pre-training gives you a model that predicts text. RL is what teaches it to reason. And here's what makes Microsoft's approach unusual: they started the RL climb from zero reasoning traces. No warm-starting from DeepSeek's thinking traces, no distillation from o1 or Claude. T he model had never seen a chain-of-thought. They had to teach it to think from scratch. They ran three parallel RL climbs on three specialist models: 1. STEM specialist: math, physics, chemistry, competitive coding 2. Agentic specialist: code execution, tool use, real-world software tasks 3. Helpfulness & Safety specialist: instruction following, human preference, refusal calibration Each specialist uses the same RL recipe but different reward signals. The RL algorithm: GRPO with two key fixes The base algorithm is GRPO (Group Relative Policy Optimization). For each problem, sample 128 responses, score them, compute relative advantages, update the model. But vanilla GRPO has two failure modes at scale: Entropy collapse - the model becomes overconfident, outputs the same answer pattern every time, and stops exploring. They fixed this with adaptive entropy control: a simple feedback controller that monitors the model's output entropy and dynamically widens or tightens the trust-region clip bounds to keep it in a target range. When entropy drops too low, the controller loosens the leash. When entropy runs too high (chaotic outputs), it tightens. No explicit entropy bonus term needed. Gradient explosions - the unclipped branches of the GRPO objective (where the new policy moves in the "right" direction) can sometimes produce catastrophic gradient norm spikes. They added a hard outer ratio clip on top of the standard trust-region clip. Two layers of protection instead of one. The reward function Three components: 1. Task reward: did the answer verify correctly? (SymPy for math, test-case execution for code, AI judge for everything else) 2. Language consistency reward: the model would start hallucinating foreign-language tokens inside its chain-of-thought as context lengths grew, which correlated with training instability. Penalize non-English words in the CoT. 3. Length penalty: adaptive by problem difficulty. Hard problems (low pass rate) get a weak length penalty - let the model think as long as it needs. Easy problems get a strong length penalty - stop the hedging and just answer. Self-distillation: the key to surviving Running RL for thousands of steps on a single continuous run is fragile. Numerical drift accumulates. Infrastructure fails. The base model gets updated. They solved all three with the same mechanism: self-distillation. Periodically, they'd collect the best rollouts from the current RL run, run SFT on a fresh mid-trained checkpoint using those rollouts, and restart the RL climb from the new checkpoint. The key finding: you need ~1M traces, sampled from later stages of the climb (not early checkpoints), with diversity of prompts mattering more than quantity per prompt. Random sampling of traces outperformed every clever selection strategy they tried. This also gave them a way to swap in a better base model mid-climb without losing months of RL progress - just distill the current RL model's knowledge into the new base, then keep climbing. The final merge: one unified model After the three specialist RL climbs complete, they do a simple SFT on a fresh consolidated model using traces from all three specialists. This gives one model that's good at STEM, agentic coding, and being helpful and safe. Then one final lightweight RL climb on that consolidated model produces MAI-Thinking-1. The result: 97% on AIME 2025. 52.8% on SWE-Bench Pro. 87.7% on LiveCodeBench v6. From a model trained entirely from scratch, zero knowledge borrowed from any other model's outputs. The pipeline is called the "hill-climbing machine." The name is literal - not a one-shot training run, but a system designed to keep improving. The interesting part isn't any single component. It's that every piece (the data pipeline, the RL recipe, self-distillation, the reward design) is built to sustain a long climb rather than peak fast and plateau.

Microsoft AI @MicrosoftAI

2 days ago

Seven new models launching at Build: let’s go! Reasoning. Code. Image. Transcribe. Voice. Built from scratch on a clean data lineage, designed for efficiency, working seamlessly as a family of models Thread 🧵 #MSBuild

131 515 3K 345K 1K

Just dropped a Linkedin Learning course recently: linkedin.com/learning/pract…

All UI will become AI This is THE article you'll need give your agents a frontend. Agent-user collaboration is the future, see below 👇

Shubham Saboo @Saboo_Shubham_

17 hours ago

x.com/i/article/2057…

26 91 607 319K 2K

CCNA 2.0 REVEALED for 2027 Cisco just announced the massive CCNA 2.0 blueprint update coming in February 2027! Discover what this completely refreshed, job-ready certification means for your IT career. Thank you to @Cisco for sponsoring this video and my trip to Cisco Live!

This is because AI agents, on Windows, trained on Stack Overflow, kept trying to use coreutils commands over and over, failed, burned tokens to use PowerShell that it didn't know. 🔗github.com/microsoft/core…

Uber reportedly now caps coding agents at $1,500/month per employee per tool - seems sensible to me, but it's also an interesting hint at the value Uber thinks these tools are providing simonwillison.net/2026/Jun/3/ube…

How OpenAI Built Its Data Agent Most teams building data agents stack routers, fine-tunes, and complex retrieval pipelines on top of multiple LLMs. OpenAI didn't. Their data agent runs on a single model and only 13 tools, across 1.5 exabytes and 90,000 tables. It's "pretty vanilla" by design. We spoke with Emma Tang, Head of Data Platform Engineering at OpenAI, to better understand the architecture and the engineering decisions behind it. The article covers: - The architecture behind the data agent - The six layers of context that make a single LLM reliable across 90,000 tables - How OpenAI Uses Codex Internally: 3 Use Cases - Five practical lessons for any team building a domain agent - Where OpenAI's data platform is headed next

Everything to know about RTX Spark in less than 90 seconds: 🟢 Personal AI agents 🟢 Faster creator workflows 🟢 RTX ON gaming All in thin and light laptops, and SFF desktops.

At Microsoft Build 2026, we’re introducing new ways for developers to learn and demonstrate skills in the tools they use every day. Today, we’re announcing: • Microsoft Pro Badge (private preview), a first-of-its-kind credential that sets a new standard for recognizing real-world skills and verified proficiency demonstrated through everyday work. • Co-authored skilling experiences with Anthropic in AI Skills Navigator. Learn in flow. Build across stacks. Get recognized for what you do. Then build further at AI Skills Fest, earn certification opportunities* and compete in an AI hackathon*. *Terms and conditions apply Learn more: msft.it/6019vjx25