ossa-ma

Where is the actual financial modelling? This is pure speculation?

I understand being bearish and frightened of AI but this accounts for absolutely NOTHING, and especially doesn't include any projections on potential ad revenue which is likely going to be huge given their DAU and what you can extrapolate their ARPU to be based on other big tech advertisers.

Langfuse is by far the best of the langs (-chain, -graph, -smith, -flow) in terms of UI/DX/integration/docs/quality.

Interesting development, announced at the same time as ClickHouse raising $400M Series D at $15B valuation.

Doesn't seem like much is changing at Langfuse - still open source, was already hosted on ClickHouse cloud, lot of collaboration existed between both teams.

Excellent advice for the 'House of Cards' politics of big tech, but it’s essentially corporate pacifism.

In any other setting you can't afford to watch money and motivation burn just to stay 'politically solvent'.

(Lalit is very good at fitting complex corporate dynamics in a single blog post though.)

Posting this because the majority of the entries are not your typical holiday destinations.

Now how much more inelastic do we get if an investor sells $1 worth of bonds and buys $1 worth of call options?

Actually its worse, those are all offer letters scraped from the internet. So Interview Coder is blatantly lying:

Apple (2017) - https://www.scribd.com/document/500958913/offer-1

Amazon - https://www.scribd.com/document/819970502/672915686-Amazon-O...

Microsoft - https://www.scribd.com/document/786362890/Arjun-Chakraborty-...

Every startup is at the mercy of the big 3 (OpenAI, Anthropic, Google).

They can and most likely will release something that vaporises the thin moat you have built around their product.

This feels like the first time in tech where there are more startups/products being subsumed (agar.io style) than being created.

That entire website may very well be AI generated including the blog post itself.

We need a new movement "Prompt Posts", instead of generating an entire blog post, share the prompt instead and save us having to skim 1000 words of AI content.

Absolutely will do so more often.

Usually I'll find something interesting and write a post about it myself but gotcha.

Long read but worth it.

a16z paved the path for a large chunk of the industry even if engineers don't like to admit it.

Even something like this would have been impossible to realise unless you had an engineering background and were hands-on:

> Ben didn’t think Hadoop was going to be the winning architecture. It was notoriously difficult to program and manage, and Ben thought it was poorly suited for the future: every step in a MapReduce computation wrote intermediate results to disk, which made it painfully slow for iterative workloads like machine learning.

Like he said a lot of OSS works on tiered features, most is free, premium at a premium.

But the main issue is that doesn't work anymore if LLMs have been trained entirely on the free and can generate the premium with ease.

When skiing in Alps, everyone is there to ski and enjoy themselves. When skiing in the US, people see it as an opportunity to use a hobby to gain an advantage, to network, to talk business, to flaunt, to compare, to use an escape as a entrance.

Same analogy can be applied to the beach, sports events, playing tennis, especially golf.

A lot of American hobby infrastructure is built around catering to the business type, why else would a ski lodge need a boardroom? Or a gym need a coworking space?

Just added. You saved me from a lot of potential legal trouble!

I completely agree, they've invested so much in AR/XR that if this doesn't work, there is zero hope for exponential growth.

Unsure on whether engineering-focused investment style posts are allowed on HN.

Thought I’d give it a go and try writing something different. Let me know what you think.

The irony of this being completely written by Claude.

When you’ve worked with Claude intimately you instantly recognise it’s writing:

- littered with “It’s not X, it’s Y.”

- Short. Punchy. Triplets.

- Overexaggeration.

A decade of reinforcement and agentic learning was spent playing games (Google Deepmind AlphaGo, AlphaStar, OpenAI Five), including against each other. So what makes it a new frontier?

Good post, outlines what a GOOD senior/staff engineer does well and identifying which of these LLMs are missing.

> But the model may need to actually do wetwork (e.g., talk to humans) to get all of the information it needs.

Even this line shows how limited LLMs are in that they have a single source of truth and guidance (the user), I've yet to see any examples of agents that are really collaborative and take multiple inputs from multiple users (not sources, users, we know LLMs can web search and file search). I think that may be the next step before an embodied independent agent.

I've adjusted the statistic for you:

Adjusted beef consumption: 4.5 million litres of water can be used to produce 300kg of beef -> US (highest beef consumer/capita) consumes 23.3kg of beef , enough to feed ~13 Americans (30 Brits, ~43 Japanese) yummy delicious grass-fed beef yearly!

And the other stats you ignored? As well as the main point I was making?

"The University of Rhode Island based its report on its estimates that producing a medium-length, 1,000-token GPT-5 response can consume up to 40 watt-hours (Wh) of electricity, with an average just over 18.35 Wh, up from 2.12 Wh for GPT-4. This was higher than all other tested models, except for OpenAI's o3 (25.35 Wh) and Deepseek's R1 (20.90 Wh)."

https://www.tomshardware.com/tech-industry/artificial-intell...

https://app.powerbi.com/view?r=eyJrIjoiZjVmOTI0MmMtY2U2Mi00Z...

https://blog.samaltman.com/the-gentle-singularity

So he's spent $51k on tokens in 3 months to build what exactly? Tools to enable you to spend more money on tokens?

Quick math on the environmental impact of this assuming 18.35Wh/1000 tokens:

Total energy: 4.73GWh, equivalent of powering 450 average US homes annually

Carbon footprint: ~1822 metric tons of CO2, equivalent of driving 4.56 million miles in a gas powered car

Water consumption: 4.5 million litres, recommended daily water intake for 4000 people for a full year

Yet they're on twitter bragging...

https://x.com/steipete/status/2004675874499842535

The beginning perfectly embodies the culture in Silicon Valley and touches on a crucial part that I notice when I visit: the complete lack of self expression or as I would put it ZERO drip.

Remove the tech, what does SF contribute to the world wrt culture? Especially when compared to other metropolitan cities: NY, London, LA, Tokyo.

I like to read and review what was captured in .inbox.md before it is committed and synced across my knowledge base. Allows me to catch mistakes, tweak preferences, add context and decide whether something is actually worth pushing.

I will typically make multiple '/handoff's per day as I use Claude code whereas I typically use '/sync' at the end of the day to organise them all at once.

There are a quadrillion startups (mem0, langmem, zep, supermemory), open source repos (claude-mem, beads), and tools that do this.

My approach is literally just a top-level, local, git version controlled memory system with 3 commands:

- /handoff - End of session, capture into an inbox.md

- /sync - Route inbox.md to custom organised markdown files

- /engineering (or /projects, /tasks, /research) - Load context into next session

I didn't want a database or an MCP server or embeddings or auto-indexing when I can build something frictionless that works with git and markdown.

Repo: https://github.com/ossa-ma/double (just published it publicly but its about the idea imo)

Writeup: https://ossa-ma.github.io/blog/double

Sorry responded to the wrong person!

LMAO you got me

Very interesting would love to read about this, do you have a link to the article?

You're right but my understanding is that Groq's LPU architecture makes it inference-only in practice.

Like Groq's chips only have 230MB of SRAM per chip vs 80GB on an H100, training is memory hungry as you need to hold model weights + gradients + optimizer states + intermediate activations.

Part of it is the irrational feeling of a disgruntled investor, doubt that I'm alone.

The other part is he has a track record of dumping on retail then telling them not to buy his next deal once he's already cashed out.

No shade but most other coverage will focus on whether this signals an AI bubble. That's missing the story.

Nvidia explicitly did NOT acquire Groq. They licensed the IP and hired the talent. This structure dodges CFIUS review (Groq had $1.5B in Saudi government contracts), antitrust scrutiny, and years of regulatory delays.

The $13B premium over the September valuation was the cost of regulatory arbitrage. Announced Christmas Eve while Trump's AI Czar (Chamath's All-In podcast co-host) is in office. Chamath's Social Capital made AT LEAST ~$2B on this exit.

My article breaks down: what Nvidia actually bought vs what they left behind, why the deal structure matters, who got paid, and the political connections nobody's talking about.