Wiredframe Radar logo

Category intelligence

Social Media Briefing — June 25, 2026

456 current items analyzed and ranked.

Executive synthesis

Social Media Summary

OpenAI's custom silicon dominated the day. The company and president Greg Brockman unveiled Jalapeño, a from-scratch LLM-inference chip built with Broadcom over nine months, signaling full-stack vertical integration. The flagship announcement drew massive reach (3.8M views), with newsletters and commentators amplifying claims of strong inference performance powering ChatGPT, Codex, and agentic workloads.

Key Themes

OpenAI Custom Silicon (Jalapeño) · 4GPT-5.5 Instant Update · 2AI Agents and Computer Use · 10Open Models & Policy · 3Agentic Tools and Future of Work · 7Open Models and Lab Economics · 5Agent Memory & Tooling · 4AI Bubble Skepticism · 5Benchmarks & Evaluation · 2Robotics & Datasets · 4

Primary evidence

Top Ranked Signals

Social Twitter Jun 24

We’ve designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and bro...

By @OpenAI

85 score
AI Analysis

OpenAI announces Jalapeño, its first in-house AI chip built with Broadcom and purpose-built for LLM workloads powering ChatGPT, Codex, the API, and agentic products, framing it as expanding its full-stack platform.

We’ve designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products. Chips are foundational to the AI economy. Building our own expands our full-stack platform from products to models to infrastructure, and will help us scale intelligence, serve more people, and expand access to AI.
AI hardwareOpenAIinference chipsAI infrastructure
Social Twitter Jun 24

There are a few things that I look back on as my mistakes in the early days. Quake was overly ambit...

By @ID_AA_Carmack

78 score
AI Analysis

John Carmack reflects on early-career mistakes during Quake development, including over-ambition, pushing teams too hard at startup intensity, and poor founder stock arrangements.

There are a few things that I look back on as my mistakes in the early days. Quake was overly ambitious technically. We could have done all the great multiplayer and modding work inside a Doom++ engine, allowing the designers to work with a more stable base instead of rug-pulling everything out from underneath them a couple times. The follow up game could have then brought in full 6DOF environments and characters. I pushed everyone too hard. I didn’t appreciate how maturing companies need more
engineering leadershipcareer lessonsstartups
Social Twitter Jun 24

Introducing Jalapeño — designed from scratch for LLM inference over nine months, accelerated by our ...

By @gdb

75 score
AI Analysis

Brockman introduces Jalapeño, an inference chip built from scratch over nine months and accelerated by OpenAI's own models, touting strong performance per watt.

Introducing Jalapeño — designed from scratch for LLM inference over nine months, accelerated by our models. Perf per watt looking incredible.
AI hardwareOpenAIinference chips
Social Twitter Jun 24

Another quick lecture -- I've been asked many times for prereq's to my book and what you should know...

By @natolambert

70 score
AI Analysis

Nathan Lambert shares a prerequisites lecture for his book covering language model basics, post-training masks, cross-entropy, KL divergence, and RL framing, built with GLM 5.2.

Another quick lecture -- I've been asked many times for prereq's to my book and what you should know, so built a little lecture (with GLM 5.2) to cover some more basics. Topics include: 00:00 Introduction & Course Prerequisites 01:37 Language Models Overview 02:47 The LM Head 04:29 Softmax & Log-Probabilities 06:13 Anatomy of an LM Training Example 06:37 Computing LLM Probabilities (+Phoebe the Dog) 09:52 Three Common Masks in Post-Training 11:03 A Small Decoding Review 12:14 Training an LM: C
AI educationML fundamentalspost-trainingreinforcement learning
Social Twitter Jun 24

Add more wins for GLM. The model has some brittle characteristics, and is getting crushed by closed...

By @natolambert

68 score
AI Analysis

Nathan Lambert praises GLM model wins while noting its brittle, jagged characteristics versus closed models, congratulating Zhipu AI (Zai_org).

Add more wins for GLM. The model has some brittle characteristics, and is getting crushed by closed models here, but we should expect open models to be more jagged, and you use multiple of them depending on the task. Congrats again to @Zai_org and am excited for the next one
open modelsGLMmodel evaluation
Social Twitter Jun 24

🧠LangSmith Engine as Sleep Time Compute Memory for agents is often described as “sleep time compute...

By @hwchase17

66 score
AI Analysis

Hwchase17 explains a LangSmith Engine tutorial implementing sleep-time compute for agent memory by tracing trajectories, analyzing them, and updating a memory store via Context Hub.

🧠LangSmith Engine as Sleep Time Compute Memory for agents is often described as “sleep time compute” or “dreaming” This involves running a background process to analyze agent trajectories and update a memory store Today, we show how to do that in LangSmith: 1. Trace all agent trajectories to LangSmith 2. Use LangSmith Engine to analyze and suggest memory updates 3. Store and update memory in Context Hub Tutorial: t.co/sFEtTUueLO
agentsmemoryLLM toolingLangSmith
Social Twitter Jun 24

We benchmarked Mistral OCR against other frontier and open-weight models on ParseBench 📊 For a mode...

By @jerryjliu0

64 score
AI Analysis

Jerry Liu reports benchmarking Mistral OCR on ParseBench, noting it is competitive for its price point with strengths in semantic formatting and weaknesses in tables and charts.

We benchmarked Mistral OCR against other frontier and open-weight models on ParseBench 📊 For a model at its price point, it is quite competitive!
  • It wins on semantic formatting - understanding strikethroughs, superscripts/subscripts, title hierarchy, links
  • It is competitive on content faithfulness (reading order + hallucinations + omissions) and visual grounding (bounding boxes)
  • It does ok on tables and doesn't really have chart capabilities.
Of course, some of the frontier models + O
OCRbenchmarksMistraldocument AI
Social Twitter Jun 24

“The Al bubble is by far the biggest stock-market bubble that mankind has ever seen - both in scale ...

By @GaryMarcus

62 score
AI Analysis

Marcus amplifies a Bloomberg quote calling the AI surge the largest stock-market bubble in history given its scale, leverage, and exuberance.

“The Al bubble is by far the biggest stock-market bubble that mankind has ever seen - both in scale and scope, in terms of the leverage that's in there and in terms of the sheer excessive exuberance that's driven valuations" - Paul Gambles via Bloomberg. t.co/jsMtDrLPMX
AI bubblemarket valuation
Social Twitter Jun 24

@salomon_diei The basic idea is easy and v0 is a hackathon project. The product here is a lot closer...

By @karpathy

62 score
AI Analysis

Continuing the discussion from yesterday, Karpathy explains his org-level harness is a genuinely different, enterprise-grade way of working, deeply integrated and multiplayer, where the system writes most code and everyone becomes a manager, distinct from Slack bots or OpenClaw.

@salomon_diei The basic idea is easy and v0 is a hackathon project. The product here is a lot closer to *it actually works*, for enterprise grade deployments, and after quite a bit of internal experimentation and iteration. It’s kind of hard to describe other than (per the post) it’s writing majority of code, it’s deeply integrated, multiplayer, and it starts to feel like everyone is a manager. So I understand it looks easy to dismiss on quick reading but it’s not some LLM Q&A with RAG over Slac
agentic toolsfuture of workdeveloper tooling
Social Twitter Jun 24

The rise of MoE models introduced new challenges in training, and @huggingface's Transformers v5 bro...

By @NVIDIAAI

62 score
AI Analysis

NVIDIA AI describes NeMo AutoModel building on Hugging Face Transformers v5 to support MoE training with Expert Parallelism, DeepEP, and TransformerEngine kernels, claiming 3.4 to 3.7x higher training throughput.

The rise of MoE models introduced new challenges in training, and @huggingface's Transformers v5 brought first-class support for solving them. Now, NeMo AutoModel builds on top of v5. Part of the NeMo framework for building models at scale, NeMo AutoModel brings optimizations to a broad set of model families through support for Expert Parallelism, DeepEP, and TransformerEngine kernels with a few lines of code. We found NeMo AutoModel brings a 3.4 to 3.7x higher training throughput for popular
NVIDIA ecosystemMoE trainingtraining infrastructureopen source tooling
Social Twitter Jun 24

The architecture of this new world model is one of the most interesting things I've seen lately: Le...

By @svpino

62 score
AI Analysis

Svpino explains the novel architecture of Tripo AI's Project Eden world model, which builds a persistent world map first then renders from it, contrasting with frame-by-frame world models that forget object states.

The architecture of this new world model is one of the most interesting things I've seen lately: Let me first explain how most world models work: They predict and render one frame at a time. If you are navigating in one of these worlds, and you look left, the model draws whatever looks right in the moment. Every time you change your viewpoint, the model has to imagine what should be there again, so it's very common for these models to "forget" what's in the world. For example, if you put a to
world modelsgenerative AI3D