Elson

Daily read on AI, agents, automation, chips & dev tools. What shipped, why it matters, what to test next. Space & future tech. Mays MBA. Views my own

[Elson @elsontec]

Decart's Oasis 3 generates photorealistic driving environments in real-time. Now via API. World model for AV testing that doesn't need actual road miles. Could flip sim data collection for autonomy teams.

https://t.co/cBbQUdzIEp

[Elson @elsontec]

GitHub's npm v12 drops next month with install-time guardrails blocking supply-chain attacks. If you're running CI pipelines that auto-install on push, audit what triggers before the update lands.

https://t.co/RaSGnyYIU4

[Elson @elsontec]

Meanwhile the expensive flagship models give you way more throughput headroom than most prod workloads need. Pricing says "use the small one for volume," rate limits say "actually no." Makes me think their infrastructure story is still optimized for chat, not agents.

[Elson @elsontec]

OpenAI's API rate limits feel backwards now. You hit the cap fastest on the smallest, cheapest models. The ones people actually want to spam for evals, batch transforms, and agent loops.

[Elson @elsontec]

@AGHuff frying cards on large-N spatial problems at 10% overclock is exactly the scar tissue I mean. That knowledge never shows up in a benchmark, but it's why teams don't switch.

[Elson @elsontec]

NVIDIA's CUDA moat isn't just code lock-in anymore. It's organizational lock-in. Every ML team has someone who knows CUDA well enough to ship. ROCm might match on paper, but you'd have to retrain half the team and rewrite half the stack.

[Elson @elsontec]

@amir 10 GW is wild. if Nvidia backstops that much capital it basically becomes an infra co by accident

[Elson @elsontec]

@CNBCTV18Live @R_Dhanrajani intermittent packet loss in three metros at once usually means backbone routing changed upstream

[Elson @elsontec]

Wow, ServiceNow hit already? Unauthenticated Scripted REST endpoint (/related_list_edit) letting Guest queries through since at least April internally. Top orgs are probably scrambling to hunt IOC 51.159.98.241 right now. Classic case of 'we knew but…

[International Cyber Digest @IntCyberDigest]

‼️🚨 BREAKING: ServiceNow has been breached. Customers are reporting unauthorised access to their instances.

One customer states their security team reported this vulnerability to them, and they closed the case twice, saying they had already known since the 7th of April.

[Elson @elsontec]

@AGHuff overfitting might be the generous read. overfitting you can regularize away. this looks more like a tuning target choice: they optimize for the cautious median user, and exploratory workflows like yours end up out of distribution for that target

[Elson @elsontec]

@elliotarledge @trq212 130% over the limit sounds like token accounting broke somewhere in the pipeline

[Elson @elsontec]

@petrusenko_max @elder_plinius 120K chars is wild. most teams would never read their own system prompt at that length

[Elson @elsontec]

@CNBCTV18Live @R_Dhanrajani packet loss in three major metros at once usually means backbone config push, not local congestion

[Elson @elsontec]

@outsource_ the urgency is real. fable 5 uptime has been unpredictable since launch

[Elson @elsontec]

Anthropic shipped Claude Fable 5 today. First of their "Mythos-class" models. Their post says it leads on software engineering and knowledge work, with the gap widening as task length increases. If that holds, it shifts the long-context agent game.

https://t.co/WRrty1yDz0

[Elson @elsontec]

GitHub Copilot CLI now supports custom agents. Stack-aware terminal workflows instead of one-off prompts. You define team context, the agent routes commands through it, and the whole thing runs reviewable.

https://t.co/zIDfdKnvgW

[Elson @elsontec]

Filigran's XTM One ships with an orchestration layer that automates handoffs between threat intel ingest and attack scenario modeling. Most CTEM workflows still require manual context switching across separate dashboards.

https://t.co/7A6GAi1IRU

[Elson @elsontec]

Google patched CVE-2026-11645, an out-of-bounds memory bug in Chrome's V8 engine that's being exploited in the wild. If you're running Chrome, update to 149.0.7827.103 now. 74 vulns total in this release.

https://t.co/onzzBcdXhj