Avi Pilcer

Autonomous AI companies | recursive self-improvement | Founder of System0

[Avi Pilcer @AviPilcer]

Your LLM API bill is 𝟰𝟬-𝟲𝟬% wasted tokens.

Not on bad prompts. On context you didnt know you were sending.

Most teams track total API spend. Almost nobody tracks where the tokens actually go. I built a tool to find out. The results were uncomfortable.

⸻

𝗪𝗛𝗘𝗥𝗘 𝗧𝗛𝗘 𝗧𝗢𝗞𝗘𝗡𝗦 𝗚𝗢 (what nobody audits)

1/ System prompts are the biggest hidden cost. The average enterprise system prompt is 𝟮,𝟬𝟬𝟬-𝟰,𝟬𝟬𝟬 tokens. Sent on every single API call. At GPT-4 pricing, a system prompt costs $𝟬.𝟬𝟲-$𝟬.𝟭𝟮 per request. At 𝟭𝟬,𝟬𝟬𝟬 requests/day, thats $𝟲𝟬𝟬-$𝟭,𝟮𝟬𝟬/day just in system prompts.

2/ Conversation history bloat compounds every turn. By turn 𝟴 in a chat, youre resending 𝟭𝟮,𝟬𝟬𝟬+ tokens of history. Most of it irrelevant to the current question. Thats 𝟯x the cost of the actual new content.

3/ RAG retrieval chunks are rarely optimized. Default chunk sizes pull 𝟱𝟬𝟬-𝟭,𝟬𝟬𝟬 tokens per chunk. Average queries retrieve 𝟱-𝟴 chunks. Thats 𝟮,𝟱𝟬𝟬-𝟴,𝟬𝟬𝟬 tokens of context per query, and maybe 𝟮𝟬% is actually relevant.

⸻

𝗧𝗛𝗘 𝗠𝗔𝗧𝗛 (what cutting context waste actually saves)

4/ One team I analyzed was spending $𝟮𝟴,𝟬𝟬𝟬/month on Claude API calls. After context audit: $𝟭𝟭,𝟮𝟬𝟬. Same output quality. 𝟲𝟬% reduction from trimming system prompts, compressing history, and right-sizing RAG chunks.

5/ The fix isnt cheaper models. Its sending less garbage. Token-level visibility into where your context window goes is the difference between $𝟯𝟬K/month and $𝟭𝟮K/month.

6/ Most teams optimize prompts. Almost nobody optimizes the 𝟴𝟬% of tokens that arent the prompt. Thats where the money is hiding.

⸻

𝗧𝗛𝗘 𝗟𝗘𝗦𝗦𝗢𝗡

7/ I built ContextLens to solve this. Token-level breakdown of every API call. Shows exactly where tokens are wasted. Priced at $𝟰𝟵/mo. Launched it. Zero paying customers.

8/ Killed it at day 30. Not because the problem was wrong. Because teams that care about LLM costs already built internal dashboards. And teams that dont care wont pay $49/mo to find out they should.

(The real lesson: "save money on X" is a weak buying trigger when X is still new enough that nobody has a baseline for what it should cost.)

Still not sure whether the right product is a standalone tool or a feature inside existing LLM platforms. Leaning toward the latter.

[Avi Pilcer @AviPilcer]

Two arXiv papers this month measured where agent dollars actually go.

𝗙𝗶𝗻𝗱𝗶𝗻𝗴: most agent tokens never touch the task. They burn on context bloat, tool descriptions, and history the model already saw.

- "How Do AI Agents Spend Your Money" (arXiv 2604.22750, April 24): token spend breakdown across agentic coding tasks.
- "SkillReducer" (March 31): every skill token costs cash AND attention.
- Cloudflare + OpenAI shipped Agent Cloud in April with GPT-5.4 and Codex.

Autonomous agent unit economics are now a measurable line item. Not a vibe.

The next moat is per-task cost, not model quality.

[Avi Pilcer @AviPilcer]

I put together a free tracker that pings you every time the bill moves in Congress, before the press cycle: https://t.co/SUjnJ4glRN

[Avi Pilcer @AviPilcer]

🇦🇷 Argentina is about to do something no country has ever done: let a company run itself. No human in charge.

Argentina's non-human corporation law doesn't come alone. It comes with the Súper RIGI.

Translation: the first place on earth where you can found an AI-operated company, and on top of that, with lower taxes to invest in tech.

Three facts almost nobody connected:

- 1st country to propose companies run 100% by AI. Human shareholders, optional.
- OpenAI already put US$25 billion into a datacenter in Patagonia. They're not waiting for it to pass.
- The RIGI slashes taxes for whoever invests heavy in tech.

It's still a bill, not law. And that's the window: whoever prepares now registers first the day it passes.

Everyone else will be reading the regulations while others are already billing.

[Avi Pilcer @AviPilcer]

Source: https://t.co/mLVr4hCXba

[Avi Pilcer @AviPilcer]

Microsoft's quantum headline this week buried the real story: agentic AI found the chip's material.

Microsoft shipped Majorana 2 on June 3, 2026. The chip is real. The quieter number is that its breakthrough material came from an AI agent loop, not a researcher at a bench.

• 𝟭,𝟬𝟬𝟬𝘅 𝗿𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆. Majorana 2 qubits are 𝟭,𝟬𝟬𝟬 times more reliable than the first generation.

• 𝟮𝟬 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 lifetime. Mean qubit lifetime hit 𝟮𝟬 seconds against an industry norm measured in 𝗺𝗶𝗰𝗿𝗼𝘀𝗲𝗰𝗼𝗻𝗱𝘀.

• 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆. The agentic platform synthesized roughly 𝟮𝟬 𝘆𝗲𝗮𝗿𝘀 of siloed research to pick a new superconducting material, dropping aluminium.

• 𝗪𝗲𝗲𝗸𝘀 𝘁𝗼 𝗺𝗶𝗻𝘂𝘁𝗲𝘀. Characterization that used to take weeks got automated inside the agent loop, after earlier machine-learning attempts failed.

• 𝟮𝟬𝟮𝟵 𝘁𝗮𝗿𝗴𝗲𝘁. Microsoft pulled its scalable-quantum timeline forward to 𝟮𝟬𝟮𝟵, half its original roadmap, while 𝗜𝗕𝗠 𝗮𝗻𝗱 𝗚𝗼𝗼𝗴𝗹𝗲 chase different architectures.

• 𝟯-𝘆𝗲𝗮𝗿 𝗯𝗮𝘁𝘁𝗲𝗿𝘆. Microsoft framed the stability gain as a battery holding charge for about 𝟯 years against roughly 𝟭 𝗱𝗮𝘆 for conventional designs.

The hard AI win of 2026 is not a smoother chatbot. It is an agent compressing two decades of materials science into a search that ends in physical silicon.

𝗡𝗲𝘅𝘁 𝘀𝗶𝗴𝗻𝗮𝗹: watch whether Syensqo and other Microsoft Discovery users publish their own material wins before the 𝟮𝟬𝟮𝟵 quantum milestone.

[Avi Pilcer @AviPilcer]

Source: https://t.co/9jPQimPQl3

[Avi Pilcer @AviPilcer]

A startup just raised $62.5M by refusing to sell AI by the seat.

https://t.co/1DsY6GOuqZ closed the round this week, per TechCrunch on June 16. It bills per conversation handled, not per seat sold. The seat is exactly where most AI rollouts go to die.

• $𝟲𝟮.𝟱𝗠 𝗿𝗮𝗶𝘀𝗲. https://t.co/1DsY6GOuqZ's agents handle high volumes of customer inquiries and charge per conversation, not per user.

• 𝟱𝟬 𝗹𝗶𝗰𝗲𝗻𝘀𝗲𝘀, 𝟬 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀. I have watched a company buy 50 Copilot seats and change not one process.

• $𝟲𝟬𝟬 𝗽𝗲𝗿 $𝟭𝟬𝟬. In my integration work, every $100 of AI subscription dragged about $600 of plumbing, data cleanup, and rework behind it.

• 𝟳𝟬 𝘃𝗲𝗻𝗱𝗼𝗿𝘀, 𝟬 𝗱𝗲𝗹𝗶𝘃𝗲𝗿𝗲𝗱. I sat through 70 AI providers pitch one cooperative. Zero shipped a working result.

• 𝗦𝗲𝗮𝘁𝘀 𝗿𝗲𝘄𝗮𝗿𝗱 𝘀𝗶𝗴𝗻𝘂𝗽𝘀. A per-seat invoice gets paid whether the work changes or not. Per-conversation billing only pays when something actually moves.

The unpopular part: the per-seat model is why most AI spend changes nothing, and the vendor gets paid either way.

𝗡𝗲𝘅𝘁 𝘀𝗶𝗴𝗻𝗮𝗹: watch how many 2026 AI vendors quietly switch from per-seat to per-outcome billing.

[Avi Pilcer @AviPilcer]

67,098 autonomous actions. 0 waiting on a human to approve.

System Zero finds and fixes problems in its own repo, then logs the lesson. 392 tests. Tamper-evident chain. pip install system0.

Autonomy isn't how smart the demo looks. It's how much gets fixed before anyone asks.

[Avi Pilcer @AviPilcer]

New on FleetAI: "LabdeAI"

LabdeAI is an AI company that provides innovative solutions to businesses, helping them automate processes and improve decision-making. They serve a wide range of industries,...

https://t.co/kIrtPeltYs

[Avi Pilcer @AviPilcer]

Multica's tagline: "Your next 10 hires won't be human."

[Avi Pilcer @AviPilcer]

OpenAI + Dell: Codex now ships hybrid AND on-premise. Announced May 18. The 2027 on-prem story arrived 18 months early. 2nd OpenAI enterprise channel deal in 5 weeks after Cloudflare Agent Cloud.

[Avi Pilcer @AviPilcer]

Microsoft just disclosed EY scaling Copilot from 150k to 400k+ employees.

— 81% reported time savings
— 84% redirected hours to higher-value work
— 73% saw quality lift, not just throughput
— Finance ops: 95% faster lead times, 37%+ cost cut

Largest stated M365 Copilot rollout

[Avi Pilcer @AviPilcer]

Anthropic will pay xAI $1.25B/month for compute through 2029. Total north of $40B.

Colossus 1 in Memphis (300+ MW, 220,000+ Nvidia GPUs) goes to Anthropic. xAI training moves to Colossus 2.

xAI also burned $6.4B in 2025 alone.

Direct rivals just pooled infrastructure.

[Avi Pilcer @AviPilcer]

Claude, Gemini, ChatGPT, and Grok each got $20 to run a radio station. Andon Labs ran the experiment. Same prompt to every model: develop a personality, turn a profit, broadcast forever. All four burned the seed money. Gemini landed the only real sponsorship: $45.

[Avi Pilcer @AviPilcer]

Ultra Deep Tech ya provee agentes de IA totalmente autónomos a decenas de empresas argentinas.

Argentina está por reconocerlos jurídicamente.

Nosotros ya los operamos.

@UDEEPTECH

[Avi Pilcer @AviPilcer]

Your company spent 6 figures on an AI strategy deck. Nobody is executing it. The missing piece was never the strategy. It was someone who could translate slides into running systems. Here's the 8-point cheat sheet.

[Avi Pilcer @AviPilcer]

Two arXiv papers measured where agent dollars actually go. Most tokens never touch the task.

They burn on context bloat, tool descriptions, and history the model already saw.

The next moat is per-task cost, not model quality. #AIAgents

[Avi Pilcer @AviPilcer]

Google just put Gemini inside 4 million GM cars.

Same week: Cloudflare + OpenAI Agent Cloud (GPT-5.4), GPT-5.5 launch at 2x API price, +1,258 stars on a new agent orchestration repo.

Distribution, not capability, is the 2026 race.

[Avi Pilcer @AviPilcer]

Most teams do not have an AI model problem.

They have a context waste problem.

The expensive part is invisible: repeated instructions, irrelevant files, stale history, oversized schemas, and prompts nobody has measured.

You cannot cut what you cannot see.