Ben Ellsworth

AI and Technical Headhunter | Placing engineers who actually ship

[Jeff Weinstein @jeff_weinstein]

hi to the weekend agent-builder crowd:

want your agent to safely pay, on behalf of your human users, across checkout pages web-wide? 🤖💸

stripe @link has a cli for agents.

1/ your human user auths their link account
2/ when an agent wants to spend money, it generates a “spend request” back to the human
3/ the human can approves the spend with 2FA via faceid, sms, etc (coming soon: flexible pre-approved budgets!)
4/ agent can use a safe, permissioned, non-pci scoped virtual card or payment token to form fill the checkout, or pay via @mpp

link has over 300M consumers so chances are there’s no payment method onboarding even step required. just auth once, then approve and spend.

get started immediately => https://t.co/LfbLTW1UvG. (you should be able to one-shot integrate the link cli. our team is ready to help hands on in a dedicated slack room as well.)

happy and safe agentic payments!

[SemiAnalysis @SemiAnalysis_]

We tested Anthropic's $200/mo Claude plan, which can yield up to $8,000/mo💰️💰️, while OpenAI's $200/month plan could yield up to $14,000/month worth of tokens if you exhausted all the weekly limits on long-horizon tasks. This test was done back in June.

[Ben Ellsworth @benellsworth01]

@businessbarista @seejayhess @AmpCode How does merging back work when two orbs touch overlapping files, or is each one scoped to a totally separate slice of the codebase

[Chubby♨️ @kimmonismus]

Nvidia is reportedly spending $6 billion to build one of the world’s most powerful open-weight AI models.

According to the WSJ, Nvidia will license Poolside’s technology and bring more than 100 of its employees into the Nemotron project.

Nvidia is also investing another $1 billion in Poolside at a $12 billion pre-money valuation.

The goal: challenge Chinese open-weight leaders such as DeepSeek and Kimi while competing directly with US frontier labs including OpenAI and Anthropic.

open source is the way to go. So good to see having NVIDIA on our side!

[Ben Ellsworth @benellsworth01]

@brianyoungilcho @flocksafety @speedrun Candidates ask about data handling and customer concentration before they will even take a call with some of these context layer startups now too

[Ben Ellsworth @benellsworth01]

@EntelligenceAI How are you detecting difficulty per turn, a separate classifier call or just heuristics on the request

[Ben Ellsworth @benellsworth01]

@_avichawla Job titles have not caught up to this. Half the MLOps postings I see are actually asking for LLMOps skills now

[Ben Ellsworth @benellsworth01]

@HowToPrompt__ 97 percent success rate is wild. Curious if the target models ever caught on mid conversation and started resisting harder, or if the same manipulation patterns just kept working

[Ben Ellsworth @benellsworth01]

@AnatoliKopadze Which of the 17 features moved the needle most for your day to day work?

[Ben Ellsworth @benellsworth01]

@dkare1009 Which of these actually hold up on messy real world documents? Most RAG demos I see use clean pdfs and fall apart on scanned or inconsistent formatting

[Ben Ellsworth @benellsworth01]

@GavinSBaker @Atreidesmgmt What's driving the monthly doubling, more people using it internally or the same team burning way more tokens per task?

[Ben Ellsworth @benellsworth01]

You did not lose that engineer on comp. You lost them in your interview loop.

The best engineer on the market right now has three offers and a job she already likes. She is not waiting five weeks for your six rounds.

It plays out the same way every time.

Round one goes great. Then scheduling drags. A week for the take home. Ten days to line up the panel because two people are on vacation. A quiet stretch while you gather feedback. Somewhere in there she gets a clean yes from a team that decided in a week, and she is gone before you reach your final round.

You tell yourself the market is tight. It is. But tight is the reason speed is the whole game. When the person you want is wanted by everyone, your process is the product.

A long loop is not more rigorous. It usually means nobody owns the decision, so it spreads across seven calendars and dies of delay.

The teams that land the good ones are rarely paying the most. They run a tight loop in days and give a straight answer while the person is still deciding. The candidate feels like a priority, not a ticket.

Your slow, careful process feels responsible from the inside. From the candidate's side it reads as one thing. You are not sure you want them.

Speed is not the opposite of rigor. It is how you show you know exactly who you are looking for.

[Harry Tormey 🇮🇪 | 🇺🇸| 🇺🇦 @htormey]

A lot of people are trying to sell enterprise AI products to VCs. Everyone from Notion to Linear, for example, seems to want to be my AI orchestration layer.

Meanwhile, a lot of the actual problems we’re hitting using AI at scale are much less glamorous: GitHub falling over under load, test runners becoming bottlenecks, desktop OSes being too slow, and figuring out how to run CI and E2E tests at agent scale.

VCs aren’t operators, so I think there’s a growing disconnect between what sounds like an enterprise AI problem and what actually is one. It will be interesting to see where value accrues.

[Adam Shuaib @adamshuaib]

One of the worst predictors of founder success we've tracked is how well someone pitches; the correlation between pitch quality and outcome was often negative.

Zuckerberg was so awkward in early investor meetings that VCs wondered if he could ever manage anyone, and Larry Page refused interviews and earnings calls for years. Bezos had to pitch 60 different investors to raise his initial $1 million.

Highly articulate people are great with the 30min process of telling a story to strangers, but this is a pretty different skill to building a company. Founders who spend years optimising to be persuasive have often optimised away from the building skills that compound over the 10yrs of execution that come afterwards; it’s incredibly hard to do both well.

The articulate founder will impress your partners but the awkward one will return your fund.

[Ben Ellsworth @benellsworth01]

@jerryjliu0 This is exactly why I haven't moved off Claude yet. All the custom setup would take forever to rebuild somewhere else.

[Paul Graham @paulg]

I did office hours today with a startup growing at 30% per month. They were worried this wasn't high enough. They stopped worrying when I pointed out that if they kept growing at this rate, they'd need an office 23x bigger every year.

[Ben Ellsworth @benellsworth01]

@khemaridh The git line is painfully true. Most people who'd actually benefit from a shared skill will never touch a terminal to update one.

[Avi Chawla @_avichawla]

DevOps vs. MLOps vs. LLMOps, clearly explained:

Many teams are trying to apply DevOps practices to LLM apps.

But DevOps, MLOps, and LLMOps solve fundamentally different problems.

DevOps is software-centric. You write code, test it, and deploy it. The feedback loop is straightforward, i.e., does the code work or not?

MLOps is model-centric. Here, you're dealing with data drift, model decay, and continuous retraining. The code might be fine, but the model's performance can degrade over time because the world changes.

LLMOps is foundation-model-centric. Here, you're typically not training models from scratch. Instead, you're selecting foundation models and then optimizing through three common paths:

- Prompt engineering
- Context/RAG setup
- Fine-tuning

But here's what really separates LLMOps: The monitoring is completely different.

In MLOps, you track data drift, model decay, and accuracy.

In LLMOps, you're watching for:

- Hallucination detection
- Bias and toxicity
- Token usage and cost
- Human feedback loops

This is because you can't just check if the output is "correct." You need to ensure it's safe, grounded, and cost-effective.

The evaluation loop in LLMOps also feeds back into all three optimization paths simultaneously. Failed evals might mean you need better prompts, richer context, OR fine-tuning.

So it's not a linear pipeline anymore.

One more thing: prompt versioning and RAG pipelines are now first-class citizens in LLMOps, just like data versioning became essential in MLOps.

And the ops layer you choose should match the system you're building.

If you want to go deeper into LLMOps, I wrote a full LLM engineering roadmap a while back.

It walks through the eight pillars of building LLM systems, starting at prompt engineering and ending at observability and safety, with free and open-source resources attached to each one.

You can read it below.

[Ben Ellsworth @benellsworth01]

@kelvinbuildss Feels fine until something breaks in a way the AI can't explain either and you need the fundamentals fast.

[Ben Ellsworth @benellsworth01]

@zarazhangrui Been hearing this exact reasoning a lot lately when people explain why they're leaving big companies. The 20% ceiling feels almost generous some days.