Huda Bilal

Founder @ 🤹🏽‍♀️🥅. FinEng/OR @uoftengineering. Prev @blackrock, @CIBCInnovation, @CancerCare_ON, @eligibleapi, @PublicisSapient 🇨🇦

[Gergely Orosz @GergelyOrosz]

That building new and complex software is not all that different today with these models vs pre-AI

Most work going into aligning ppl, making a plan that makes sense for the business and everyone else, coordinating w partners, getting architecture right hard at first, rework happens after a pilot etc

[Nate @natebjones]

I keep thinking about agents as a test of how honestly a company understands its own work.

If the process only works because a few experienced people quietly correct the system all day, then the workflow was never really encoded.

It was being subsidized by judgment.

[Séb Krier @sebkrier]

I think there's a slow, growing realization that "the model will fix it all, and if not the next one will, so all you can do is to wait, there's no way to compete" is bullshit.

It seems obvious to me that there are, and will continue to be, many opportunities in building companies that can leverage ("scaffold") AI models to address all sorts of problems in sectors and contexts that they know best.

This is efficient even in the face of growing capabilities, for the same reason as it remains more cost effective for a model to call on a calculator rather than to try to resolve the problem internally.

Some reasons for this are proprietary data flows, regulatory needs, deep integration with crusty legacy systems, customer trust earned over time, and tacit knowledge. So to be clear this is not a story about thin wrapper around an API - shallow/fake specialization gets eaten ofc.

People focus on labs because that's sexy, and to be clear they'll continue to be, and I don't think you'll see thousands of companies investing that kind of capex; but I think a big part of the story (and value creation) will look like the nameless businesses that never get discussed on social media, by journalists or commentators.

This is great news for entrepreneurship and shaking up incumbents, and bad news if your view of the world is "the eye of Sauron will simply ingest everything, all that is left is the zero sum fight for power and escaping the permanent underclass".

On the other hand I expect lots of bubbly dynamics and failing businesses and questionable commercial ventures (always has been etc), partly because the territory is still so uncertain that many can still make a quick buck while VCs are in (temporary) Opus psychosis mode. But this too shall pass.

[James Rosen-Birch ⚖️🕊️ @provisionalidea]

I have a lot of *thoughts* on the Dorsey piece, but tonight I’ll just reiterate — while I very much love that people are starting to care about org design again,

1) human context is not the same thing as LLM context, and environments of ubiquitous surveillance and documentation do not miraculously transmute one into the other

2) reinventing the flat org for the umpteenth time will not magically make its flaws go away just because you plug in AI. I know tech perennially fantasizes about flat orgs and ‘firing all the managers’ once every two to three years, but there are much cooler and more impactful ways to redesign your org around AI that actually optimize for what the AI’s good at as opposed to trying to force AI to resurrect an undead fantasy

3) strategy, planning, resource delegation, coordination, assignment, advocacy, conflict resolution, mentorship, accountability, and decision-making under uncertainty (collectively: management) makes up a distinct skillset and area of expertise that becomes *more* valuable in an agentic world, not less. when you give every engineer ten agents to assign work to, what you’ve done is turned those engineers into managers of digital workers. this then *increases* the administrative and coordination burden of the org geometrically despite headcount remaining stable.

4) the exciting thing about AI in this moment is that it can empower people to make faster, better-informed decisions, not that you can hand your decisions off to a machine to make them in your stead.

[Hugo Navarro @HugoNavarroPer2]

EBITDA (2026 edition): Earnings Before Iran, Tariffs, and Donald Announcements.

[James Rosen-Birch ⚖️🕊️ @provisionalidea]

I think it’s about time to point out that the meaning of “context” for LLMs (reference text for next token prediction) is fundamentally different from the meaning of “context” for people and organizational systems (history, relationships, tacit knowledge, prior experience and learned intuition), and the growing homonymic conflation of the two is becoming significant enough to be deeply counterproductive and harmful to the successful implementation of AI in the workplace

capturing and recording tacit knowledge is a very difficult problem. applying it effectively is equally challenging, and the basis of expertise. these things cannot be captured and replicated by ‘simply’ implementing ubiquitous workplace surveillance and feeding all emails and meeting transcripts into a language model.

the proper solution is to design the flow and structure of work such that the machines do tasks machines are good at (search, retrieval, summarization, rote assembly and stitching), and people do the integrative and creative thinking, planning, learning, and analysis, augmented and bolstered by the machines.

[James Rosen-Birch ⚖️🕊️ @provisionalidea]

no. good research, useful data, wrong diagnosis and conclusion.

a bit related to some long-term research of mine, but —

this is not an “AI” problem

this is a “cognitive load, interface, and form factor” (ie cognitive ergonomics) problem

it should be absolutely no surprise to anyone that when AI companies prioritize, design for, and market “fire and forget”, “full autonomy”, and “let the machine make the decisions so you don’t have to”, the result is that people take the path of least resistance and use the tool in the way it is designed to be used.

this does not mean it is the only way to use AI, that it is an inevitable outcome of LLMs, or that this outcome is a foregone conclusion.

[no earthquake (undignified) @no_earthquake]

LLMs have convincingly demonstrated that coding is the easiest activity, maths is medium hard, and having taste is the hardest

[James Rosen-Birch ⚖️🕊️ @provisionalidea]

there’s been a lot of quiet, careful work over the past year between the major tech companies (mainly Google, but also OAI) and financial infrastructure giants (Visa, Mastercard) to build out the plumbing for agent-to-agent commercial transactions and prepare the market for it

[Jake @JustJake]

When you become an engineer in Canada, you take an oath

To never cause harm/suffering/etc to humans as a result of your work

This applies equally, if not more, to business

Both can be used for good (to remove suffering) or evil (to add it)

Do good work. Don't be evil

[François Chollet @fchollet]

The idea that we will automate work by building artificial versions of ourselves to do exactly the things we were previously doing, rather than redesigning our old workflows to make the most out of existing automation technology, has a distinct “mechanical horse” flavor

[James Rosen-Birch ⚖️🕊️ @provisionalidea]

In 2002, an NIST study said software bugs cost the US economy $59B/year ($106B in today’s dollars). It was 0.6% of US GDP at the time.

Three years ago (2022), a CISQ study estimated software bugs now cost the US economy $2.41 trillion, or 10% of US GDP.

[evan conrad @evanjconrad]

If there is a bubble in AI, the following describes where the bubble is.

Investors, founders, and team members in Silicon Valley are bad about conceptualizing inventory & credit risk and many put no premium onto lowering it.

In software, the only real marginal input is cloud spend. However, in a functioning software business, you end up with high margins that improve over time. Every customer you bring on grows your revenue at a faster rate than your AWS bill increases.

However, in AI, GPU spend grows proportionally to your revenue, because of the bitter lesson. Every time you click generate, it actually uses more GPUs. This causes a ripple effect throughout the value chain, because everyone, everywhere is price sensitive. That compresses margins for GPU owners, who end up in a price war with each other. That forces them with a choice: (1) sell short-term contracts which expose them to the risk that the customer churns and they’re still paying their loan or (2) sell long-term contracts for lower margins which ship the risk problem to the customer. The better option is 2.

But their customers are the inference layers, who now are saddled with a long-term contract on compute and are now faced with the same problem as their vendors: do you sell short term contracts to your customers (usage based pricing at the end of the month) or do you sell long-term contracts? But this, of course, shoves the problem off to the application layer!

So now you’re at the application layer, which is now being forced into either paying long-term contracts but is primarily selling short-term or usage based services for low margin.

And there’s the bubble. If the application layer falters, then the whole thing unravels. The application layer defaults on their contracts, it means the revenue for the inference layer isn’t coming in, which means the revenue for the cloud layer isn’t either, and the debt machines that fund the clusters get blown up. Since the scale of the money going into compute is large, it causes investors to rebalance their portfolios out of high-risk assets, drying up venture capital, making more of the application layer to falter.

The saving grace here is the companies in the application layer are making money (lots of it) and the growth really is there. This stuff works and is going to continue to work.

However, it’s likely worthwhile for companies that are used to thinking in terms of software to start thinking about this. Application layers that don’t have long-term contracts of their own should avoid signing long-term compute or inference deals if they can’t get liquidity and wise investors should discourage their portcos them from doing so.

When you walk into SF Compute's office, there's a sign that just says "risk is neither created nor destroyed". The only way to resolve the problem is to price the risk & diffuse it.

[Nabeel Hyatt @nabeel]

Low trust environments engender funky deal structures, which is why real estate people love a funky deal.

Tech has trended the opposite for 40 years, simpler deals, higher trust, make the pie bigger. That is, until the last few years.

I'm not commenting on a specific deal, but funky deal structures are not a good trend. Putting lawyers and deal boyz in the front seat instead of builders is bad.

And it has long term effects on mantaining a high trust environment in tech which we all benefit from, which frankly is being assaulted in multiple directions the last year or two.

[Huda Bilal @DraftingPencil]

What real leadership is made of.

[William Allen @williamallen]

I don’t think he’s ever told the story, but it’s worth telling. When we were selling @Behance to Adobe many years ago, @scottbelsky made a spreadsheet of every employee (32 of us at the time) and personally negotiated each persons title, salary and incentive structure, and made that part of the overall deal terms. I heard the phone calls where he went to bat for each individual.

He not only didn’t have to do this, but it actually complicated some of the other factors in the deal. It changed the trajectory of so many people’s lives, including my own.

2 years later, 100% of that original team was still at Adobe. Even today, a dozen years later, many of the core members are still there, building.

I was inspired by it then, and I'm inspired by it now.

[ayman nadeem @aymannadeem]

I’m a YC/VC-backed founder.

building is hard enough without tech cheering on open racism.

saying vile things about muslims + arabs is now seen as “edgy” in VC Twitter. it’s not. it’s just bigotry.

if you’re a muslim/swana founder trying to build something good, DM me. let’s not feel alone in this mess.

[will depue @willdepue]

its crazy how Slack has obliterated the corporate landscape by massively empowering the naturally socially anxious/autistic to be absolute demons in text channels and destroyed the average high EQ slow-typing product guy who just did not grow up fighting flame wars on Discord

[Tyler Angert @tylerangert]

most product management tools are actually just anxiety management

[Raven Kwok @RavenKwok]

An interactive Apollonian gasket.