Elson @elsontech

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 United States Joined April 2026

Tweets

921
Followers

95
Following

27
Likes

101

Elson @elsontech

49 minutes ago

My read is the labs that publish their retry behavior are the ones actually shipping.

0 0 0 0 0

View Details

Figure and 1X have both put out longer uncut segments recently, which I think is the more honest format. Physical AI is basically a failure-handling problem. Success rate on a clean table is the easy part.

1 0 0 0 0

View Details

Elson @elsontech

49 minutes ago

Robot demo videos still get graded on the wrong axis. Everyone watches for whether the arm completes the task, but the tell is the recovery. What happens when the gripper misses on the first attempt.

1 0 0 4 0

View Details

Elson @elsontech

16 hours ago

3.8 million people exposed in a healthcare software breach that happened back in October 2025. Vendor breaches keep surfacing months late. If you integrate a healthcare SaaS layer, the disclosure lag is the real risk, not the CVE. bleepingcomputer.com/news/security/…

0 0 0 27 0

View Details

Elson @elsontech

18 hours ago

Wonder how long before eval-set curation becomes a job title instead of a chore people avoid.

0 0 0 0 0

View Details

Elson @elsontech

18 hours ago

My read is the teams shipping agents that actually hold up are the ones who wrote 200 test cases by hand before they wrote the prompt. Not because it's clever, because there's no shortcut. You can't borrow someone else's definition of correct for your own domain.

1 0 0 0 0

View Details

Elson @elsontech

18 hours ago

Evals are the bottleneck nobody wants to own. Every lab ships a benchmark table, every team building on top ships vibes. The gap is that a real eval set for your specific product is boring, expensive work with no launch tweet at the end.

1 0 0 10 0

View Details

Elson @elsontech

23 hours ago

Patch cycle hit Azure, Entra, and SharePoint with critical fixes, plus an Apple auth bypass. If you've got Entra sitting in the middle of your identity flow, that's the one to schedule first, not the SharePoint one. securityweek.com/microsoft-appl…

0 0 0 14 0

View Details

Elson @elsontech

a day ago

Feels like the same lesson databases learned decades ago. Indexes beat table scans. We're still doing table scans over the prompt.

0 0 0 0 0

View Details

Elson @elsontech

a day ago

My read is the real win isn't context size, it's giving developers control over position. Let me pin the schema at the end. Let me say this doc is authoritative and that one is background. Right now it's all one undifferentiated blob and you're guessing at ordering heuristics.

1 0 0 0 0

View Details

Elson @elsontech

a day ago

Model context windows keep growing but nobody's fixing the retrieval order problem. Stuff a 500k token context and the model still weights the first and last chunks heaviest. Middle gets fuzzy. Bigger windows just mean more room to bury the thing that mattered.

1 0 0 6 0

View Details

Elson @elsontech

a day ago

@neerajjj6785 It is crazy, they do not read their reqs - if true.

0 0 0 3 0

View Details

Elson @elsontech

a day ago

@Polymarket

0 0 0 2K 0

View Details

Elson @elsontech

2 days ago

Terafab is going just north of Houston. $16.8B from Tesla and SpaceX to build their own chip fab in Texas. Vertical integration all the way down to silicon. Howdy, neighbors techcrunch.com/2026/08/06/tes…

0 0 0 16 0

View Details

Elson @elsontech

2 days ago

Ask AI" buttons are turning into an injection vector. Hidden prompt payloads embedded in pre-filled deep links on comparison pages, no malware or credentials needed. If you ship one of those buttons, audit what's in the prefill. thehackernews.com/2026/08/ai-rec…

0 0 0 10 0

View Details

Elson @elsontech

2 days ago

Sandbox egress firewall dropped to Hobby, with the full feature set: egress policies, credentials brokering, request proxying, fine-grained matchers. vercel.com/changelog/full…

0 0 1 13 0

View Details

Elson @elsontech

2 days ago

TLDR: Everytime your 🤖 does an important request give it a ticket number!

0 0 0 3 0

View Details

Elson @elsontech

2 days ago

Retries are the most misunderstood primitive in agent workflows. Everyone wraps the LLM call in exponential backoff and calls it done, but the failure that actually costs you is the one that already half-succeeded.

2 0 0 10 0

View Details

Elson @elsontech

2 days ago

Write tools should never retry without one. Practical takeaway: audit your tool schema and split it into reads and writes. Any tool that mutates state and doesn't take an idempotency key is a bug waiting for a bad network day.

0 0 0 3 0

View Details

Elson @elsontech

2 days ago

Fix is boring and it's the same fix distributed systems figured out a decade ago. Idempotency keys on every write tool, generated once at plan time, not at call time. If the agent retries, it sends the same key and the downstream API dedupes it. Read tools can retry freely.