Pinned Tweet
@hackernoon's editors picked my essay as a Top Story today.
Thesis in one line: technology never destroys value, it relocates scarcity. AI made execution cheap. The price moved to judgment, verification, and a name willing to be answerable.
Original: piszczek.pl/blog/scarcity-…: .co/7TntPgu5wk
Thesis in one line: technology never destroys value, it relocates scarcity. AI made execution cheap. The price moved to judgment, verification, and a name willing to be answerable.
Original: piszczek.pl/blog/scarcity-…: .co/7TntPgu5wk
1
2
9
425
+67% more tokens per second from the same Blackwell.
Qwen3.8 27B: 35.04 → 58.53 tok/s with 70 → 120 W plus VRAM/core tuning.
The twist: Qwen3.6 35B A3B cared far more about memory speed than the last watts.
158 requests, charts, CSVs:
piszczek.pl/blog/rtx-pro-4…
Qwen3.8 27B: 35.04 → 58.53 tok/s with 70 → 120 W plus VRAM/core tuning.
The twist: Qwen3.6 35B A3B cared far more about memory speed than the last watts.
158 requests, charts, CSVs:
piszczek.pl/blog/rtx-pro-4…
2
0
0
41
Hugging Face's Carbon-X gets within 0.8 percentage points of GPT-6.1 Sol on five DNA benchmarks, at roughly 1/1800 of the reported inference cost per problem.
The averages: Carbon-X 66.7%, Sol 67.5%, Claude Opus 5.5 60.9%.
Sol and Opus ran at max reasoning. Carbon-X used no per question. It beats Sol on two tasks and trails on three: 74.9% vs. 71.0% on coding-variant effects, but 85.1% vs. 91.7% on complex variants. Two of the five benchmarks were built by the release team.
This is a joint research release with Zhongguancun Academy, Telethon and the University of Naples. The results are team-reported, and Carbon-X has not been clinically validated.
The cost ratio compares standard API pricing with measured Carbon-X throughput on an H200 at $5/hour. It does not include the full cost of research or validating the answers.
Carbon-X has 14.8B parameters and builds on Qwen3-14B with 302 billion tokens of additional training, 55% raw DNA. DNA and text enter the same decoder, with six bases per DNA token. Apache 2.0 weights fit on a single 80 GB GPU in BF16.
In a separate, smaller evaluation, requiring Kimi K3 to consult Carbon-X raised its average from 49.4% to 60.5%. Both setups had Python; the test used 100 questions per benchmark across five benchmarks. Those scores are not directly comparable with the main table. This experiment does not establish the same 1,800x workflow saving.
The cost gap makes specialization worth considering; the Kimi result shows one way a larger model can benefit.
The architecture is the interesting part: a general assistant coordinates the task, a specialist interprets the DNA, and code handles exact operations.
A smaller model can change what the larger one is for.
➕ Follow me for the model-card details that change how AI work gets divided and paid for
The averages: Carbon-X 66.7%, Sol 67.5%, Claude Opus 5.5 60.9%.
Sol and Opus ran at max reasoning. Carbon-X used no per question. It beats Sol on two tasks and trails on three: 74.9% vs. 71.0% on coding-variant effects, but 85.1% vs. 91.7% on complex variants. Two of the five benchmarks were built by the release team.
This is a joint research release with Zhongguancun Academy, Telethon and the University of Naples. The results are team-reported, and Carbon-X has not been clinically validated.
The cost ratio compares standard API pricing with measured Carbon-X throughput on an H200 at $5/hour. It does not include the full cost of research or validating the answers.
Carbon-X has 14.8B parameters and builds on Qwen3-14B with 302 billion tokens of additional training, 55% raw DNA. DNA and text enter the same decoder, with six bases per DNA token. Apache 2.0 weights fit on a single 80 GB GPU in BF16.
In a separate, smaller evaluation, requiring Kimi K3 to consult Carbon-X raised its average from 49.4% to 60.5%. Both setups had Python; the test used 100 questions per benchmark across five benchmarks. Those scores are not directly comparable with the main table. This experiment does not establish the same 1,800x workflow saving.
The cost gap makes specialization worth considering; the Kimi result shows one way a larger model can benefit.
The architecture is the interesting part: a general assistant coordinates the task, a specialist interprets the DNA, and code handles exact operations.
A smaller model can change what the larger one is for.
➕ Follow me for the model-card details that change how AI work gets divided and paid for
1
0
6
157
PHP just patched two critical vulnerabilities that can allow attackers to execute code without logging in.
The warning covers PHP-FPM, Apache mod_php, CGI, FrankenPHP and LiteSpeed.
On mod_php, the demonstrated attack runs before your application code does.
Your WordPress x the runtime underneath it.
Both vulnerabilities score 9.2 on CVSS v4.0. Here is the ecosystem behind that advisory:
→ 37.3 million live WordPress sites identified by BuiltWith.
→ PHP on 69.3% of sites whose server-side language W3Techs can identify.
→ About 5.1 million PHP-fingerprinted IPs in VirusTotal.
→ Only about 1.1 million of those IPs disclose a PHP version in headers.
These figures describe the dependency footprint, not a count of exploitable installations. Sites and IPs cannot be added. Actual exploitability depends on the deployment and attack path.
Session upload progress is enabled by default. A default can be inherited for years without anyone choosing it for that particular application.
Turning it off blocks confirmed session paths of one flaw. It does not remediate the separate out-of-memory flaw or every path of the state-lifecycle flaw. PHP says updating is the only complete solution.
-> Now the AI part.
Wiz says AI helped its researchers develop and adapt working exploits across multiple applications in hours.
Cloudflare reports attack traffic and attackers changing payloads to evade its protections. That does not establish AI involvement in those attacks.
Fast automation was already dangerous. In 2024, Cloudflare observed TeamCity exploitation attempts 22 minutes after a public proof of concept appeared.
What concerns me now is the cost of adapting an attack to the next deployment, while defenders still coordinate across hosting providers, image maintainers and application teams.
A rebuilt image can be available while an old container keeps receiving traffic. WordPress can be current while PHP remains vulnerable.
That is where the potential blast radius grows: one shared dependency, many separate patch decisions.
An attacker can reuse an approach. Defenders still have to close each exposed deployment.
➕ Follow me for the infrastructure details that determine how far an AI-era security failure can spread
The warning covers PHP-FPM, Apache mod_php, CGI, FrankenPHP and LiteSpeed.
On mod_php, the demonstrated attack runs before your application code does.
Your WordPress x the runtime underneath it.
Both vulnerabilities score 9.2 on CVSS v4.0. Here is the ecosystem behind that advisory:
→ 37.3 million live WordPress sites identified by BuiltWith.
→ PHP on 69.3% of sites whose server-side language W3Techs can identify.
→ About 5.1 million PHP-fingerprinted IPs in VirusTotal.
→ Only about 1.1 million of those IPs disclose a PHP version in headers.
These figures describe the dependency footprint, not a count of exploitable installations. Sites and IPs cannot be added. Actual exploitability depends on the deployment and attack path.
Session upload progress is enabled by default. A default can be inherited for years without anyone choosing it for that particular application.
Turning it off blocks confirmed session paths of one flaw. It does not remediate the separate out-of-memory flaw or every path of the state-lifecycle flaw. PHP says updating is the only complete solution.
-> Now the AI part.
Wiz says AI helped its researchers develop and adapt working exploits across multiple applications in hours.
Cloudflare reports attack traffic and attackers changing payloads to evade its protections. That does not establish AI involvement in those attacks.
Fast automation was already dangerous. In 2024, Cloudflare observed TeamCity exploitation attempts 22 minutes after a public proof of concept appeared.
What concerns me now is the cost of adapting an attack to the next deployment, while defenders still coordinate across hosting providers, image maintainers and application teams.
A rebuilt image can be available while an old container keeps receiving traffic. WordPress can be current while PHP remains vulnerable.
That is where the potential blast radius grows: one shared dependency, many separate patch decisions.
An attacker can reuse an approach. Defenders still have to close each exposed deployment.
➕ Follow me for the infrastructure details that determine how far an AI-era security failure can spread
0
1
3
288
DeepSeek just open-sourced its inference engine. Its own chart shows why "faster AI" needs two numbers.
On its published decode curve, more concurrent requests raise output per GPU while lowering each user's tokens per second. The host processes more work; each user waits longer output.
A provider can optimize for a very different number than its customer.
The code makes another condition inspectable. It converts measured steps per second into token rates using 1 + 3 draft tokens × 75% assumed acceptance = 3.25. That 75% is in the chart caption and passed to the plotting script, not measured by it.
DeepInfer arrived today under Apache 2.0, supporting V4.1 Flash on NVIDIA SM100 and Huawei Ascend 950. DeepSeek says the framework powers its production services. The public release provides offline inference, with additional components still needed for a complete service.
These are DeepSeek's curves on NVIDIA hardware, not a comparison with another framework or Huawei. The published code lets other operators inspect how the numbers are constructed and adapt the engine to their own workloads.
That is useful openness. It also makes the commercial question clearer: does higher throughput become a lower bill, or more waiting for the customer?
More throughput for the provider can mean less speed for you.
➕ Follow me for the trade-offs hidden inside AI performance claims
On its published decode curve, more concurrent requests raise output per GPU while lowering each user's tokens per second. The host processes more work; each user waits longer output.
A provider can optimize for a very different number than its customer.
The code makes another condition inspectable. It converts measured steps per second into token rates using 1 + 3 draft tokens × 75% assumed acceptance = 3.25. That 75% is in the chart caption and passed to the plotting script, not measured by it.
DeepInfer arrived today under Apache 2.0, supporting V4.1 Flash on NVIDIA SM100 and Huawei Ascend 950. DeepSeek says the framework powers its production services. The public release provides offline inference, with additional components still needed for a complete service.
These are DeepSeek's curves on NVIDIA hardware, not a comparison with another framework or Huawei. The published code lets other operators inspect how the numbers are constructed and adapt the engine to their own workloads.
That is useful openness. It also makes the commercial question clearer: does higher throughput become a lower bill, or more waiting for the customer?
More throughput for the provider can mean less speed for you.
➕ Follow me for the trade-offs hidden inside AI performance claims
0
0
1
44
MiniMax just benchmarked its new model inside Claude Code and Codex.
Using a lab's agent does not have to mean buying that lab's model.
M3.1 Flash launched today at $0.15 per million input tokens and $0.60 per million output tokens on the pay-as-you-go API.
The benchmark y. Terminal-Bench 4.0 used a Codex harness. CyberGym used Claude Code. DeepSWE used mini-SWE-agent. These are company-reported results.
MiniMax's current docs also show how to run M3.1 Flash inside both rival tools using its M Plan subscription. The subscription and the API tariff are separate billing routes.
That creates an awkward split for the big labs. Their agents can become the interface developers trust while another company supplies the inference.
The client still organizes tools, permissions and the working session. The model comes with its own constraints: MiniMax requires thinking to stay on, and its Codex guide recommends a 512K working context rather than the full 1M.
For a challenger, a tool developers already know is a route into an existing habit.
The agent can win the desk and lose the inference bill.
➕ Follow me for the business consequences hidden in AI launch details
Using a lab's agent does not have to mean buying that lab's model.
M3.1 Flash launched today at $0.15 per million input tokens and $0.60 per million output tokens on the pay-as-you-go API.
The benchmark y. Terminal-Bench 4.0 used a Codex harness. CyberGym used Claude Code. DeepSWE used mini-SWE-agent. These are company-reported results.
MiniMax's current docs also show how to run M3.1 Flash inside both rival tools using its M Plan subscription. The subscription and the API tariff are separate billing routes.
That creates an awkward split for the big labs. Their agents can become the interface developers trust while another company supplies the inference.
The client still organizes tools, permissions and the working session. The model comes with its own constraints: MiniMax requires thinking to stay on, and its Codex guide recommends a 512K working context rather than the full 1M.
For a challenger, a tool developers already know is a route into an existing habit.
The agent can win the desk and lose the inference bill.
➕ Follow me for the business consequences hidden in AI launch details
3
0
3
163
15 days. 1,542 agents. One goal. No follow-up prompts.
We’re living in the future. Work is being repriced.
The rich will own the agents. The poor will compete with them. Denial won’t pay the bills
We’re living in the future. Work is being repriced.
The rich will own the agents. The poor will compete with them. Denial won’t pay the bills
0
0
1
29
Anthropic says it moved its monorepo off Git as daily merges grew more than tenfold. The bigger change is where it tests the code.
The company reports about 2,500 merges a day in May and more than 30,000 in September. At that scale, it says it has not been able to replicate a e-queue lock. It invested in automated testing after merge and before deployment instead.
Seven people built Scion, its source-control system, Antique for code review, and related tools, migrating in five months. Server-side operations and sparse working copies reduce the repository data clients need. Anthropic reports a 25% CI CPU capacity gain.
Engineers still review changes for intent and steer agents away from violating system constraints. Concurrent changes can still conflict. A merge records code entering the repository; it does not establish that the code is ready for customers.
Scion fits an unusual monorepo and an internal token budget Anthropic says exceeds what other businesses would realistically spend. Copying the system is a different decision from learning from the constraint.
For a CTO, faster code generation pays when verified changes reach users sooner, without the savings disappearing into repair and rollback.
The automation dividend depends on the release gate.
➕ Follow me for AI engineering and economics, with the numbers checked
The company reports about 2,500 merges a day in May and more than 30,000 in September. At that scale, it says it has not been able to replicate a e-queue lock. It invested in automated testing after merge and before deployment instead.
Seven people built Scion, its source-control system, Antique for code review, and related tools, migrating in five months. Server-side operations and sparse working copies reduce the repository data clients need. Anthropic reports a 25% CI CPU capacity gain.
Engineers still review changes for intent and steer agents away from violating system constraints. Concurrent changes can still conflict. A merge records code entering the repository; it does not establish that the code is ready for customers.
Scion fits an unusual monorepo and an internal token budget Anthropic says exceeds what other businesses would realistically spend. Copying the system is a different decision from learning from the constraint.
For a CTO, faster code generation pays when verified changes reach users sooner, without the savings disappearing into repair and rollback.
The automation dividend depends on the release gate.
➕ Follow me for AI engineering and economics, with the numbers checked
0
0
2
53
Meta's Muse Wonder Full just overtook GPT-6 Astra on the AA Intelligence Index.
OpenAI now has another frontier rival with billions of users already in reach.
Muse Wonder Full scores 56. Claude Fable 5.1 and GPT-6 Astra score 53. Claude Opus 5.5 still leads at 58.
Two points eta brings that capability into its existing products, the commercial pressure gets much bigger than the ranking suggests.
There is a distinction worth keeping: ordinary Muse Wonder scores 52 and is available through the API. Full is available to Muse Code subscribers. Its per-token price has not been published.
Meta already has the audience. Bringing enough capability into those daily habits would raise the hurdle for a separate AI subscription.
OpenAI will have to give customers a reason to keep opening another app and paying another bill. Anthropic faces the same question. A lead on the tasks people actually need can justify that. An aggregate index cannot settle it.
For operators, the buying decision still comes down to useful work completed, and the cost of completing it.
A few benchmark points are a fragile moat against a daily habit
OpenAI now has another frontier rival with billions of users already in reach.
Muse Wonder Full scores 56. Claude Fable 5.1 and GPT-6 Astra score 53. Claude Opus 5.5 still leads at 58.
Two points eta brings that capability into its existing products, the commercial pressure gets much bigger than the ranking suggests.
There is a distinction worth keeping: ordinary Muse Wonder scores 52 and is available through the API. Full is available to Muse Code subscribers. Its per-token price has not been published.
Meta already has the audience. Bringing enough capability into those daily habits would raise the hurdle for a separate AI subscription.
OpenAI will have to give customers a reason to keep opening another app and paying another bill. Anthropic faces the same question. A lead on the tasks people actually need can justify that. An aggregate index cannot settle it.
For operators, the buying decision still comes down to useful work completed, and the cost of completing it.
A few benchmark points are a fragile moat against a daily habit
1
0
5
237
OpenAI just made 20+ announcements at DevDay.
Its new model costs a fifth of the flagship.
Its new speed costs 6x.
Read the whole list and it's one strategy.
Six theses:
1️⃣ Intelligence is a commodity. GPT-6.1 Sol: $2/$10 per 1M tokens, near-Astra by OpenAI’s numbers. 5.5 the same a day earlier. Two labs, one price tag.
2️⃣ Speed is the premium. Ultrafast: up to 8x faster for 6x the price. On Astra that’s $60/$300. Waiting became a line item.
3️⃣ Labor is nearly free. Dots are always-on agents, and the first one is included in the plan. At most 27 cents an hour. All five use cases on OpenAI’s page end with a human reviewing or approving.
4️⃣ Execution left your laptop. Codex now runs fully in the cloud. Agents API with computer use is in beta.
5️⃣ Distribution is the moat. Plugin extensions put your product inside ChatGPT’s 1.2B weekly users. Sign in with ChatGPT lets your plan allowance power 16 partner tools, Devin and Notion among them. And enterprises can now spend their OpenAI commitment on 32 partners, from Adobe to CrowdStrike. Commitments are becoming a currency.
6️⃣ The workspace is the lock-in. Spaces and Pages put teammates, ChatGPT and their dots on one shared surface. OpenAI is building its own Notion.
Now read across.
Intelligence is priced like a commodity. Everything around it is priced like a product: speed, seats, workspace, distribution, approval.
That's not a model release. That's a business model change.
Read that again.
Intelligence is becoming the loss leader.
➕ Follow me for what AI companies actually sell, one price list at a time
Its new model costs a fifth of the flagship.
Its new speed costs 6x.
Read the whole list and it's one strategy.
Six theses:
1️⃣ Intelligence is a commodity. GPT-6.1 Sol: $2/$10 per 1M tokens, near-Astra by OpenAI’s numbers. 5.5 the same a day earlier. Two labs, one price tag.
2️⃣ Speed is the premium. Ultrafast: up to 8x faster for 6x the price. On Astra that’s $60/$300. Waiting became a line item.
3️⃣ Labor is nearly free. Dots are always-on agents, and the first one is included in the plan. At most 27 cents an hour. All five use cases on OpenAI’s page end with a human reviewing or approving.
4️⃣ Execution left your laptop. Codex now runs fully in the cloud. Agents API with computer use is in beta.
5️⃣ Distribution is the moat. Plugin extensions put your product inside ChatGPT’s 1.2B weekly users. Sign in with ChatGPT lets your plan allowance power 16 partner tools, Devin and Notion among them. And enterprises can now spend their OpenAI commitment on 32 partners, from Adobe to CrowdStrike. Commitments are becoming a currency.
6️⃣ The workspace is the lock-in. Spaces and Pages put teammates, ChatGPT and their dots on one shared surface. OpenAI is building its own Notion.
Now read across.
Intelligence is priced like a commodity. Everything around it is priced like a product: speed, seats, workspace, distribution, approval.
That's not a model release. That's a business model change.
Read that again.
Intelligence is becoming the loss leader.
➕ Follow me for what AI companies actually sell, one price list at a time
0
0
4
157
Anthropic shipped a model whose most expensive setting is its worst deal.
Opus 5.5 at medium: 54.7% for ~$0.80/task
Opus 5.5 at max: 54.5% for ~$6.30/task
8x the bill, lower score.
Opus 5.5 at medium: 54.7% for ~$0.80/task
Opus 5.5 at max: 54.5% for ~$6.30/task
8x the bill, lower score.
1
0
2
71
For two years "make it better" meant "let it think longer."
That stopped being true today.
Buy the setting, not the model.
That stopped being true today.
Buy the setting, not the model.
1
0
0
9
Full breakdown of what this does to cost per accepted task: lnkd.in/p/dEErydKQ
0
0
0
11
Jensen Huang just told Dreamforce that AI needs no new laws. Same day, NVIDIA published the one law it cannot repeal.
San Francisco, September 15, on stage with Benioff: "We don't need any new laws. Safety is an engineering problem, not a legal one. You pace yourself until you Clara, same afternoon. NVIDIA's AI Infra Summit blog, on what actually paces AI:
→ Lambda ran 19 nodes inside the power budget of 16
→ Throughput 4M to 5M tokens per second, +24%
→ Performance per watt +23%, same silicon, same site
→ Vera Rubin NVL72 with DSX MaxLPS: up to 40% more GPUs per megawatt, "no new power lines required"
The regulation debate is about who gets to say slow down. Congress, the labs, or the market. The answer was already on NVIDIA's own blog. The utility says it, every day, in a unit called the megawatt. Interconnection queues run years. Chips ship in weeks.
So NVIDIA's real product this week is software that argues with the utility.
Read that again. The CEO says the industry paces itself. The engineers spent the day removing the only pacing mechanism that has ever held.
Laws regulate intent. Megawatts regulate output.
The model wars are over. The megawatt wars just started.
San Francisco, September 15, on stage with Benioff: "We don't need any new laws. Safety is an engineering problem, not a legal one. You pace yourself until you Clara, same afternoon. NVIDIA's AI Infra Summit blog, on what actually paces AI:
→ Lambda ran 19 nodes inside the power budget of 16
→ Throughput 4M to 5M tokens per second, +24%
→ Performance per watt +23%, same silicon, same site
→ Vera Rubin NVL72 with DSX MaxLPS: up to 40% more GPUs per megawatt, "no new power lines required"
The regulation debate is about who gets to say slow down. Congress, the labs, or the market. The answer was already on NVIDIA's own blog. The utility says it, every day, in a unit called the megawatt. Interconnection queues run years. Chips ship in weeks.
So NVIDIA's real product this week is software that argues with the utility.
Read that again. The CEO says the industry paces itself. The engineers spent the day removing the only pacing mechanism that has ever held.
Laws regulate intent. Megawatts regulate output.
The model wars are over. The megawatt wars just started.
0
0
1
84
Could RSI AI be the last invention we'd ever need?
"Need" is the interesting word. We'd still have curiosity, even if progress stopped needing us to invent every next step.
The broad version remains unproven. A 40-second explainer:
piszczek.pl/blog/rsi-ai-re…
"Need" is the interesting word. We'd still have curiosity, even if progress stopped needing us to invent every next step.
The broad version remains unproven. A 40-second explainer:
piszczek.pl/blog/rsi-ai-re…
0
0
1
46
Could RSI AI become humanity's last necessary invention?
People might keep inventing. Progress might simply stop waiting for our next idea.
A 1965 argument, real experiments, and a future we haven't demonstrated:
piszczek.pl/blog/rsi-ai-re…
People might keep inventing. Progress might simply stop waiting for our next idea.
A 1965 argument, real experiments, and a future we haven't demonstrated:
piszczek.pl/blog/rsi-ai-re…
0
0
2
40
Between 2004 and 2012 I broke into computer systems. It turned out to be an education in economics disguised as a technical career.
Every wall exists because something behind it is scarce. Knock the wall down and the scarcity doesn't die. It moves.
Every wall exists because something behind it is scarce. Knock the wall down and the scarcity doesn't die. It moves.
1
0
0
50
A manifesto is suspect by its own logic, so I'm paying the toll. Every claim here gets preregistered, metric and deadline, before results arrive, and published even when it humiliates me.
An opinion asks you to agree. A signal shows you where its author bleeds.
An opinion asks you to agree. A signal shows you where its author bleeds.
1
0
1
13
Full essay, with the receipts and the lineage: piszczek.pl/blog/scarcity-…
0
0
0
4