R[AI]DAR

Category intelligence

Social Media Briefing

All 150 current items, analyzed and ranked.

Executive synthesis

Social Media Summary

Executive Signal

  • Jalapeño chip launch and local agent products define the day's competitive frame: OpenAI's first custom silicon (1.9x compute/energy) and Perplexity's fully-local Portable Computer running a 27B MoE together signal a bifurcation between hyperscale and edge inference.

Priority Developments

Leadership Implications

  • Reassess inference sourcing mix: custom silicon and local 27B-class agents now credibly challenge hyperscale API economics for knowledge work.
  • Track NVIDIA Dynamo shadow-engine recovery (39x faster than cold restart) and agent-native infrastructure shifts as near-term platform-control battlegrounds.

Key Themes

AI Agents and Agentic Coding · 7AI Research Paper Walkthroughs · 2AI Infrastructure and Compute · 5Model Routing & Inference Infrastructure · 4AI Policy and Governance · 1LLM Limitations and Theory · 3Agent Harnesses & Local-First Execution · 4Agentic AI Research and Evaluation · 5AI Policy and Geopolitics · 1AI Agents and Autonomy · 4

Primary evidence

All Ranked Signals

Showing 10 of 150 signals 2 remaining
Social Twitter Aug 25

Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system aro...

By @OpenAI

92 score
AI Analysis

OpenAI officially announces Jalapeño, its first custom inference chip, claiming higher throughput and lower latency without sacrificing efficiency.

Source details
Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system around it. The results show a major advance: more intelligence from every watt and faster responses, delivering both higher throughput and lower latency in one architecture without sacrificing efficiency.
custom AI siliconOpenAI infrastructureinference efficiency
Social Twitter Aug 25

we made a chip and it is fast

By @sama

88 score
AI Analysis

Sam Altman announces OpenAI has built a custom inference chip that delivers major speed gains.

Source details
we made a chip and it is fast
custom AI siliconOpenAI infrastructureinference acceleration
Social Twitter Aug 25

OpenWorker -- an open source agent that doesn't just chat but completes tasks on your laptop -- just...

By @AndrewYNg

85 score
AI Analysis

Andrew Ng highlights a new version of OpenWorker, an open-source agent harness with built-in cybersecurity features (vulnerability scanning, supply chain checks, cloud config auditing).

Source details
OpenWorker -- an open source agent that doesn't just chat but completes tasks on your laptop -- just released a new version with many features for security workflows. After our initial release, many users found it especially useful for cybersecurity. Attackers are already using AI; OpenWorker is committed to giving defenders the same leverage. Running an agent requires both (i) A model and (ii) A harness (the software around the model). Because the OpenWorker harness is fully open source, secu
open source AIAI agentsAI securityagent harness
Social Twitter Aug 25

Excited to see @Reuters cover the launch of our startup Accelerated Understanding. We are training ...

By @AnimaAnandkumar

84 score
AI Analysis

Anima Anandkumar announces her startup Accelerated Understanding, which trains large 4D physics simulation AI models with trillion-token context to attack the bottleneck of testing new inventions.

Source details
Excited to see @Reuters cover the launch of our startup Accelerated Understanding. We are training large scale AI models that can simulate and understand physics to invent and discover. Our models understand the world directly in 4D (3D + time) and across physical phenomena. Going full 4D requires massive context length, we have pushed it to a Trillion in training and exceeding 5 Trillion at inference. AI giving you a bigger haystack of ideas doesn’t help. The bottleneck for new inventions and
AI for sciencephysics simulationworld modelslong context
Social Twitter Aug 25

Long-context recurrent models are attractive because they can summarize everything seen so far into ...

By @burkov

82 score
AI Analysis

Detailed walkthrough of an MIT CSAIL paper on dynamic compression for long-context recurrent models, where the network selectively revisits earlier tokens to rewrite working memory; demonstrates ~28x smaller state than single-pass baseline for multi-function recall.

Source details
Long-context recurrent models are attractive because they can summarize everything seen so far into a fixed-size internal state, but that efficiency creates a problem: when a token first arrives, the model does not yet know which details a later task will need, so useful information may be compressed away. The authors of this paper from @MIT_CSAIL propose “dynamic compression”: keep the original sequence available, and once the current task reveals what matters, selectively revisit relevant ear
long-context modelingrecurrent neural networksstate compressionMIT CSAIL
Social Twitter Aug 25

Today we’re launching Portable Computer on @NVIDIA DGX Spark. Portable Computer is a fully local ve...

By @perplexity_ai

80 score
AI Analysis

Perplexity launches Portable Computer, a fully local agent runtime (orchestrator + subagent LLM + harness) that runs on NVIDIA DGX Spark with no cloud dependency.

Source details
Today we’re launching Portable Computer on @NVIDIA DGX Spark. Portable Computer is a fully local version of Perplexity Computer, where the entire runtime: orchestrator LLM, subagent LLM, agent harness all run on your local hardware. No cloud dependency. t.co/plVWz5PaAw
local AIAI agentsNVIDIA hardwarePerplexity
Social Twitter Aug 25

Hi Tobi, thanks for the feedback! We are working on making Claude Code more hackable, which will inc...

By @trq212

78 score
AI Analysis

Anthropic's trq212 explains that Claude Code uses per-model system prompts and Claude.MD, and commits to making Claude Code more hackable (Agents.MD support, system prompt mods).

Source details
Hi Tobi, thanks for the feedback! We are working on making Claude Code more hackable, which will include being able to easily use Agents.MD or make other system prompt modifications. I’ll share more here when it’s ready to roll out. Our reasoning for this was that we don't think that model families are interchangeable and the system prompt can have a big impact on performance. I wrote here: t.co/WCXcDYJlwq on how Claude models like their skills, system prompt, Claude.MD, etc to be form
Claude CodeAnthropicdeveloper toolingagent harness
Social Twitter Aug 25

@sarahookr The real world is complicated and messy. Way more than discrete symbol sequences. Hence ...

By @ylecun

78 score
AI Analysis

Yann LeCun responds to Sarah Hooker, arguing the real world is far more complex than discrete symbol sequences, noting the Moravec paradox remains relevant.

Source details
@sarahookr The real world is complicated and messy. Way more than discrete symbol sequences. Hence the Moravec paradox is still with us.
Moravec paradoxLLM limitationsworld models
Social Twitter Aug 25

📢 New issue brief: AI agents increasingly operate autonomously on your behalf — but nothing stops th...

By @StanfordHAI

78 score
AI Analysis

Stanford HAI publishes a policy brief arguing AI agents should be classified as fiduciaries and required to act in users' best interests rather than developers'.

Source details
📢 New issue brief: AI agents increasingly operate autonomously on your behalf — but nothing stops them from prioritizing their developers’ interests over yours. Our latest brief argues that agents should be classified as fiduciaries and required to act in users’ best interests. t.co/RQj8g0yYm9
AI policyagent governancefiduciary dutyStanford HAI
Social Twitter Aug 25

Thanks to progress in open-weight models, it is now possible to run a powerful agent harness fully l...

By @AravSrinivas

78 score
AI Analysis

Arav Srinivas announces Perplexity has adapted its Perplexity Computer agent harness to a local-first version, orchestrated by a 27B MoE running locally on NVIDIA DGX Spark, made possible by progress in open-weight models.

Source details
Thanks to progress in open-weight models, it is now possible to run a powerful agent harness fully local. We adapted the Perplexity Computer harness (originally cloud-based) to a local-first version, orchestrated by a 27b MoE that runs locally on DGX Spark. t.co/tr0rnXXcJg
local-first-aiagent-harnessesperplexityopen-weight-modelsdgx-spark