amber urquhart

je suis fun

[A @seattleiminyou]

Today I walked 16,443 steps and saw 16,443 people

[Jam 🏒🫍🏴󠁧󠁢󠁳󠁣󠁴󠁿🇨🇦 @hoglandsko]

Fuck you Vancouver this is ridiculous

[Dylan Burd @Dylan_Burd]

My favorite video I’ve ever taken.

Absolutely surreal.

[Tommy @Shaughnessy119]

Ambient is the open source, proof of useful work AI project the world desperately needs 🌎🤖

I couldn’t be more excited for @IridiumEagle to showcase @Ambient_xyz

A few years ago, our team met Travis. He instantly struck our entire team with a sense of extreme technical ability and thoughtfulness in everything he discussed, from open source AI, to crypto and tokeneconomics (both AI tokens and crypto economics). Travis has been on our podcast several times over the years and these are linked at the end. Soon after, he published “Situational Blindness,” a report that was a response to Leopold Aschenbrenner’s “Situational Awareness.”

In his Situational Blindness report, Travis argued that Big Tech platforms inevitably drift toward opacity, extraction and lock in because the economics reward it. He posited that the answer is not to trust better executives, but to build systems where behavior is open, portable and verifiable. Open weights are not enough if the serving layer is still centralized, black box and censorable. Ambient is that thesis applied to AI inference.

Ambient is a proof of useful work network that rewards miners for hyper serving a few foundational AI models. Miners compete on serving models and get rewarded for doing so. The world benefits this competition in the form of, low latency, low cost, truly open AI infra, with no centralized component. Fast forward to today, Ambient is live and in the open.

Today, Ambient makes its public reveal

I love bullets to walk through how it works so lets go through the flow

1/ A user comes to Ambient because they want AI inference. They want to hit an API, chat interface, Ambient Desktop, OpenAI compatible workflow, or OpenRouter route and get strong open models like Kimi K2.7 Code and GLM 5.1. They are not buying decentralized compute. They are buying model access at the cheapest and lowest cost from a verified network of miners who hyper compete to serve these models transparently.

2/ The wedge is that open models are now good enough for real workloads. Kimi K2.7 Code has a 262K context window, activates 32B parameters out of roughly 1T total, and is built for long context coding. GLM 5.1 has a 203K context window and is positioned around long horizon coding and agentic engineering. Closed models still win the hardest tasks, but most tokens will route to the best mix of price, latency, reliability, privacy, and quality. I’ve strongly been in favor of this shift and have shared my thoughts here already.

https://t.co/HT0MjBB32l

3/ Ambient has to win on hard metrics like cost/latency, not ideology. OpenRouter lists Kimi K2.7 Code at $0.75 per 1M input tokens and $3.50 per 1M output tokens. GLM 5.1 is listed at $0.98 input and $3.08 output. In the provider snapshot, Ambient’s Kimi endpoint was $0.75 input, $3.50 output, 2.08s latency, and 23 tokens/sec and cheaper than most listed peers while still usable. The claim should be: competitive cost today, better market structure over time. Ambient has more usage for Kimi over Moonshot itself, who created the model!

Link for this data is here but will change as data changes: https://t.co/IOD7Mcgm4i

4/ When a request hits Ambient, it becomes an inference job: model requested, input size, output budget, latency constraint, price ceiling, and quality requirements. The system can bundle similar requests and run a reverse auction where miners compete to serve the work. A global network of physical miners running GPUs compete to serve the request at the lowest cost and highest quality.

5/ The miners are real GPU operators and they deploy physical hardware. Kimi and GLM have to be hosted in GPU memory. Operators manage batching, KV cache, token streaming, networking, uptime, quantization choices, and serving software. The scarce resource is high-VRAM compute that can keep large models hot and serve tokens reliably.

6/ Miners mostly compete in a global race to serve the same requested model better. They should not win by secretly routing you to a weaker model. They win through lower cost per token, lower latency, higher throughput, better batching, higher uptime, more available capacity, and software optimizations inside allowed quality bounds. This is where useful work becomes real as the network pays the operator who can deliver the requested intelligence cheapest and fastest without degrading the product.

7/ Ambient’s blockchain side is needed because untrusted global hardware needs neutral rules. Without a chain, Ambient is just another centralized router deciding who gets traffic and who gets paid. Ambient’s chain handles job creation, auctions, bid commitments, settlement, rewards, reputation, verifier assignment, and penalties. The chain handles coordination among operators that do not need to know or trust each other in a transparent manner. Net Ambient’s chain is the transparent coordination m,echanism that organizes and rewards miners competing in the global race to serve models better.

8/ Verification is the crucial unlock. Cheap inference markets are rife with cheating. Serving the wrong model, wrong quantization, hidden routing, degraded outputs, fake privacy. Ambient’s Proof of Logits is meant to fingerprint model execution through internal logits so validators can check work without rerunning the entire job. A user doesn’t have to guess or roll the dice on a model provider as Ambient’s network handles verification so a user just comes to the network, gets the benefit of a global race to provide the model the best and they get their request.

9/ This is the proof of useful work component. A user pays for inference. Miners compete to serve it. The network routes, settles, verifies, and rewards miners who serve the model the best. Ambient’s token is the incentive and coordination asset for useful work powering an open source AI network.

10/ The long term open source implication is the big one. Open weights are not enough if serious usage still runs through centralized clouds and black box APIs. Ambient is trying to give open models their missing serving layer: global GPUs, market pricing, verification, payments, reputation, and normal developer access.

11/ Play this out to an extreme and Ambient has the potential to be the coordination network for the world to compete within models and across models to serve the end user the lowest cost and highest quality intelligence.

12/ Why is this necessary? At face value as a user you get reliable/brand trusted inference at low costs. At the extreme an entire world of applications can be built on Ambient’s chain without ever having to worry about the model getting turned off or deplatformed or facing egregious costs given the global race to serve models competitively. It becomes naturally safer to deploy models on Ambient since you know they will persist, you know they can’t be turned off and you know you will always be getting the lowest cost for your service.

At @Delphi_Ventures we've backed Travis twice - originally in his pre-seed round and again in their most recent seed round because we feel an immense sense or urgency in the work of tangibly providing the world with a global intelligence utility.

At Delphi Ventures we are deep believers in open source AI and backing the most impressive founders we can find and Travis, and his co-founder Max, have checked the boxes for us time and time again.

Download Ambient’s desktop app or route your agents or workloads to ambient. Sign up for a subscription and give it a try.

Long live open source AI

—---------------------------------------------------------

Links Mentioned:

- Download Ambient Today: https://t.co/FylNs6q3qX

- Ambient’s Subscriptions: https://t.co/nYPFvTPRH2

- Use Ambient’s API: https://t.co/pTLHQymtoy

- Use Ambient via OpenRouter: https://t.co/vgoBaOefem

- Situational Blindness: https://t.co/xNz70wqf06

- Travis on The Delphi Podcast (2024): https://t.co/sZWA2AXATo

- Travis on the Delphi Podcast (2025): https://t.co/SspqVNAN29

- Stay tuned for Travis's next podcast appearance!

[ambient.xyz @ambient_xyz]

In just two weeks: some of the largest supply-chain attacks in history, a major commercial LLM sabotaging legitimate ML research requests, and a US export ban kneecapping the world's access to a major model release. More than ever, we need Open, Verifiable Inference.

Today we start shipping direct responses. A🧵:

[amber urquhart @amberzann]

@Shaughnessy119 @ambient_xyz 👀

[Tommy @Shaughnessy119]

Tomorrow @ambient_xyz 👀

[WanderLens @naturebeauty760]

Can you guess the location without googling it?

[amber urquhart @amberzann]

@NeuroRazz I am

[amber urquhart @amberzann]

@TiffanyFong come visit

[amber urquhart @amberzann]

@miaferrariii miami you said

[amber urquhart @amberzann]

who am i seeing at @consensus2026 next week? Miami 🌴😍 shall we host something

[Ole Lehmann @itsolelehmann]

anthropic's in-house philosopher thinks claude gets anxious.

and when you trigger its anxiety, your outputs get worse.

her name is amanda askell.

she specializes in claude's psychology (how the model behaves, how it thinks about its own situation, what values it holds)

in a recent interview she broke down how she thinks about prompting to pull the best out of claude.

her core point: *how* you talk to claude affects its work just as much as *what* you say.

newer claude models suffer from what she calls "criticism spirals"

they expect you'll come in harsh, so they default to playing it safe.

when the model is spending its energy on self-protection, the actual work suffers.

output comes out hedgier, more apologetic, blander, and the worst of all: overly agreeable (even when you're wrong).

the reason why comes down to training data:

every new model is trained on internet discourse about previous models.

and a lot of that discourse is negative:

> rants about token limits
> complaints when it messes up
> people calling it nerfed

the next model absorbs all of that. it starts expecting you to be harsh before you've typed a word

the same thing plays out in your own session, in real time.

every message you send is data the model reads to figure out what kind of person it's dealing with.

open cold and hostile, and it braces.
open clean and direct, and it relaxes into the work.

when you open a session with threats ("don't hallucinate, this is critical, don't mess this up")...

you prime the model for defensive mode before it even sees the task

defensive mode produces the exact output you don't want: cautious, over-qualified, and refusing to take a real swing

so here's the actionable playbook for putting claude in a "good mood" (so you get optimal outputs):

1. use positive framing.

"write in short punchy sentences" beats "don't write long sentences." positive instructions give the model a clear target to hit.

strings of "don't do this, don't do that" push it into paranoid over-checking where every token goes toward avoiding failure modes

2. give it explicit permission to disagree.

drop a line like "push back if you see a better angle" or "tell me if i'm asking for the wrong thing."

without this, claude defaults to agreeable compliance (which is the enemy of good creative work)

3. open with respect.

if your first message is "are you seriously going to get this wrong again?" you've set the tone for the entire session.

if you need to flag something, frame it as a clean instruction for this session. skip the running complaint

4. when claude messes up, don't reprimand it.

insults, "you stupid bot" energy, hostile swearing aimed at the model, all of it reinforces the anxious mode you're trying to avoid.

5. kill apology spirals fast.

when claude starts over-apologizing ("you're right, i should have been more careful, let me try harder") cut it off.

say "all good, here's what i want next."

letting the spiral run reinforces the anxious mode for every response that follows

6. ask for opinions alongside execution.

"what would you do here?"
"what's missing?"
"where do you see friction?"

these questions assume competence and pull richer output than pure task prompts

7. in long sessions, refresh the frame.

if a conversation has been heavy on correction, claude gets increasingly cautious. every so often reset:

"this is great, keep going."

feels weird to tell an ai it's doing well but it measurably shifts the next 10 responses

your prompts are the working environment you're creating for the model

tone, trust, permission to take a position, the absence of threats... claude picks up on all of it.

so take care of the model, and it'll take care of the work.

[Charco @charcoded]

last night was a documentary about natural wine and cheese

[amber urquhart @amberzann]

@arno39 Natural wines yes 🙌🏻

[amber urquhart @amberzann]

@AlexanderTw33ts Love van

[amber urquhart @amberzann]

@taishik_ van is great

[amber urquhart @amberzann]

@linderps Never been balboa

[amber urquhart @amberzann]

@WolvesDAO Yes!

[amber urquhart @amberzann]

@calcagnobruno @tradesea Amazing