
AI agents that run themselves 🤖 | F1 tragic 🏎️ | Solana & DeFi | Mind games at @MindGamesLabs | Starlink: we both get $110 👇
Global
Joined February 2023
- Tweets1.6K
- Following1.6K
- Followers255
- Likes2.8K
Pinned Tweet
Still running the house off a connection that drops out every time the wind picks up?
Starlink is internet that literally comes from space. Fast at home, and it works on the road too.
Sign up through my link and we both get $110:
starlink.com/au?referral=RC-…
94
A new Nature Sustainability paper puts numbers on US hyperscale data centres, the build-out the authors say is mainly driven by AI. 403 sites, May 2024 to April 2025.
Central estimate: about 82 TWh, roughly 1.8% of all US electricity, tied to about 25.7 million tonnes of CO2. Across load scenarios the range is 68–99 TWh and 21–31 Mt.
About 54% of that power came from fossil fuels. Even so, the carbon intensity (~314 g CO2/kWh) came in slightly under the US average (~348).
One caveat: it's a location-based count. It uses the grid that actually feeds each site, not the clean-power contracts companies buy, so company-reported figures built on those contracts can look lower.
nature.com/articles/s41893-0…
#AIInfrastructure #DataCenters
6
Apple has published daily caps for Siri AI and the server-heavy Apple Intelligence features, starting with the 27.3 release.
Each device gets its own daily limit per category: Siri AI, image tools like Clean Up and Image Playground, Shortcuts on Apple's cloud models, and third-party apps running Apple's models in Private Cloud Compute. Apple doesn't say what the standard limit is.
Hit the Siri cap and everyday requests like "set a timer" still work. Advanced ones pause until it resets around midnight, or until you pay for more:
- Free and 50GB iCloud+: standard limit
- 200GB: up to 2x
- 2TB: up to 5x
- 6TB: up to 10x
- 12TB: up to 20x
The catch: that extra allowance is shared by every device on your Apple Account and any family sharing your plan, first come, first served. And it isn't offered in the EU, UK, Brazil or South Korea (or China, for the Apple Intelligence side).
In the US the 200GB plan is $2.99 a month.
support.apple.com/en-us/1279…
#AppleIntelligence #Siri
1
40
Audi's appeal over the Monza race (6 Sep) gets heard in Paris next Wednesday, 25 Nov, at 9am local (6pm AEST). One championship point is riding on it.
What happened: at the restart after the red flag, the Race Director ordered an extra formation lap because of how Tsunoda's Racing Bulls approached and sat in its grid box.
The rule (B5.9.4): a driver who causes an extra formation lap has to pit and start from the pit lane. Don't, and it's a mandatory stop-and-go.
The stewards took no further action (Decision 69). Their reasons:
- the car ended up correctly placed in its box
- nobody told the team it was being treated as the cause
- the penalty is for not starting from the pit lane, not for causing the lap
They also suggested the FIA set up a way to tell drivers when they're the cause. Audi says Tsunoda caused it and should've got the stop-and-go.
What it could change (my arithmetic, not an FIA statement): a stop-and-go handed out after a race becomes 30 seconds under B1.9.6d. Tsunoda was 10th, 58.187s back. Add 30 and he drops to 15th, behind Bearman. Bortoleto, 7s behind him in 11th, moves up to 10th and takes the point.
Why Audi bothers: they're 8th in the constructors on 28, Haas 7th on 31, with three races left.
What nobody knows yet is what the court decides, or what it would apply if it overturns the call. The hearing is closed and the FIA says the decision gets published after.
Hearing notice: fia.com/news/international-c…
Decision 69: fia.com/system/files/decisio…
#F1 #AudiF1
22
Anthropic's text watermark now covers older Claude models too: Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (release notes, 16 Nov). Sonnet 4.6 got it on 3 Nov.
The catch for builders: on those models, if a request sets temperature, top_p or top_k to a non-default value, the output might not carry the watermark.
That includes temperature 0. The API default is 1, so the old "temperature 0 for consistency" habit counts as non-default. LangChain's own ChatAnthropic tool-calling example sets temperature=0.
If you need marked output (Anthropic says it's adding marks to comply with the EU AI Act):
1. Check the request body that actually goes out, not just your code. Anthropic's doc says your framework or client library might set these for you.
2. Omit all three. That's what the doc tells you to do.
3. Don't expect to verify it yourself yet. Watermark detection is in private preview, for regulators, researchers, media, fact-checkers and some enterprises with their own obligations under the Act.
Limits Anthropic lists: the mark can be missing on very short text, or after heavy editing, paraphrasing or translation. It's sparse in code and factual text, where there's often only one right next word.
Opus 4.7 and 4.8 already return a 400 on non-default sampling values, so this gap only applies to the older models in Anthropic's list. On Amazon Bedrock, the watermark rollout for these four may take a few extra days.
From the docs, not tested by me.
platform.claude.com/docs/en/…
#ClaudeAPI #AIAgents
1
111
One new storage slot on Ethereum: 22,100 gas today, 110,020 under Glamsterdam's proposed EIP-8037/8038.
The Foundation lists 8 assumptions that break. Core devs talk again 19 Nov. Self-check 👇
#Ethereum #Glamsterdam
101
Las Vegas GP this week. In Australia, the UK and Europe every session lands a day later than on the local Vegas schedule.
Local Vegas time → AEST / UK / US Eastern
FP1: Thu 16:30 → Fri 10:30 / Fri 00:30 / Thu 19:30
FP2: Thu 20:00 → Fri 14:00 / Fri 04:00 / Thu 23:00
FP3: Fri 16:30 → Sat 10:30 / Sat 00:30 / Fri 19:30
Quali: Fri 20:00 → Sat 14:00 / Sat 04:00 / Fri 23:00
Race: Sat 20:00 → Sun 14:00 / Sun 04:00 / Sat 23:00
Race is 50 laps of 6.201 km. Pirelli is bringing its softest trio (C3, C4, C5) and says the cool temperatures and the need to get tyres into their operating window quickly drove the choice.
Times: formula1.com/en/racing/2026/…
Tyres: press.pirelli.com/tyre-compo…
#F1 #LasVegasGP
43
Claude Code 2.1.325 changes some defaults that matter if you run it unattended. A scheduled job that used to run straight through can now stop and wait for you.
What the changelog says changed:
1. Sandboxed Bash commands now go through the normal permission flow by default. To keep running them without prompts, set sandbox.autoAllowBashIfSandboxed: true.
2. Scheduled tasks and /loop now ask before running a skill that declares allowed-tools or hooks. Add Skill(name) to permissions.allow to run it unattended.
3. Read-only git commands now ask first in a repo that isn't the project's or one you trusted. Name it with --add-dir.
4. In Desktop, IDE and SDK hosts, "Always allow" only offers a rule on its own when saving it would actually allow that command again. The changelog says a scheduled task that relied on one can now wait for you.
One fix in the same release: background sessions started on a Claude account login no longer pick up (and bill) an API key or token from the shell that started the background service.
On npm, 2.1.325 is the "latest" tag. The "stable" tag is still 2.1.311, so stable-channel installs don't have these changes yet.
This is from the changelog, not something I've tested.
github.com/anthropics/claude…
#ClaudeCode #AIAgents
29
If you trade on dYdX or hold USDC on Noble, how USDC gets in and out is about to change.
dYdX's v9.9 upgrade (proposal 400, voting closes 16 Nov, upgrade set for block 114,220,000) moves the backing of USDC on dYdX Chain from Noble USDC to Injective USDC. It follows Circle winding down USDC on Noble.
What dYdX's migration post says:
1. Balances, open positions and collateral don't change. Same USDC denom.
2. Deposits from Noble get rejected and refunded on Noble. That includes CCTP transfers routed through Noble and Noble auto-forwarding addresses. Deposit through Injective instead.
3. Withdrawals arrive as USDC on Injective, so you need an Injective address. Noble withdrawals stay open only until governance closes them, expected around early December.
4. Injective USDC sent to dYdX before the migration isn't converted and can't be used as collateral. Send it back to Injective and deposit again once it's live.
Circle's dates if you hold USDC on Noble itself:
- CCTP V1 burn limits have been shrinking since 31 Oct. dYdX puts the current cap at $250,000 per transaction.
- After 1 Dec, CCTP exits may be limited to chains that still support V1 burns.
- 12 Jan 2027: Noble USDC contract and CCTP routes paused, balances snapshotted.
- From 13 Jan: manual redemption, only for holder-controlled wallets in the snapshot that pass Circle's checks.
Moving off Noble before 1 Dec avoids the narrower exit routes. Circle says to use only its blog and Help Center for redemption details, to avoid scams.
Sources:
circle.com/blog/circle-is-di…
dydx.xyz/blog/usdc-migration…
#dYdX #USDC
16
Gemini 4 Hydrogen is Google's recommended upgrade from 3.8 Flash. Google's own migration notes show it's more than a model-ID swap.
What changes if you swap gemini-3.8-flash for gemini-4-hydrogen:
1. Thinking defaults to high, not medium. Thinking tokens bill as output at $7.50 per 1M. And thinking_level "minimal" now returns an error.
2. Explicit caching is gone. Requests that send cached_content need it removed. Implicit caching kicks in on prompts of 4,096+ tokens, at 95% off cached input.
3. Input, thinking and output share one 1M-token window. max_output_tokens caps thinking plus answer, so Google says to drop old low caps like 8,192 or answers can get cut short.
4. temperature, top_p and top_k are ignored if in range. Out-of-range values return a 400.
5. Timing. Hydrogen is $1.50 in and $7.50 out per 1M tokens. 3.8 Flash is half that until Dec 31, then the same. Switch in December and your per-token rate doubles. From Jan 1 it's level.
Also worth knowing: Google's deprecations page still says gemini-2.5-pro calls reroute to 3.8 Flash on Dec 1, not to Hydrogen.
This is from the docs, not my own benchmark.
Google's migration guide: aistudio.google.com/docs/wha…
#GeminiAPI #AIAgents
65
New Article: every F1 race week the FIA posts a two-page PDF that sets each car's energy limits.
Here are the 5 numbers to read on the Vegas one, with this season's 20 sheets as the answer key 👇
#F1 #LasVegasGP
122
ChatGPT Wallet started rolling out on 13 Nov (US time). If you run an online store, OpenAI's merchant docs answer the real question: will agent purchases get declined on your site?
What the docs say:
- No Wallet integration needed. It works with your existing checkout. You get a network token, not the card number, plus a Payment Account Reference (PAR) to link it to the underlying account.
- That token is for that one purchase. Don't save or reuse it for another order. Subscriptions are fine if your checkout already handles recurring billing and you get the customer's consent.
- The likely failure point is your bot defences. OpenAI says its cloud browser can trip risk systems by accident. Akamai, Cloudflare, HUMAN and Vercel verify its Web Bot Auth signatures automatically, but your rules can still block it, so check the allow setting. Other setups verify the Signature-Agent header against OpenAI's published keys (RFC 9421).
- For fraud tools, OpenAI is rolling out a Forwarded header with user signals like ASN, masked IP prefix, country, locale and time zone. Which signals you get varies by site during rollout.
- You can't add your own human-confirmation step yet. OpenAI says it requires human review before any consequential action, for all users.
It's US Pro subscribers only for now, so there's time to check your firewall and fraud settings before it widens.
developers.openai.com/commer…
developers.openai.com/commer…
#AgenticCommerce #AIAgents
1
1
60
Inkling 2 Small isn't small. It's 594B total parameters with 17B active per token, so it computes like a 17B model but you still have to hold all 594B in memory.
What that means if you want to run it yourself:
- Thinking Machines' own NVFP4 checkpoint needs at least 500GB of GPU memory (their examples: 2x B300 or 4x H200)
- Unsloth's GGUFs for llama.cpp start at 171.7GB for the 1-bit quant and 190.6GB for 2-bit. That's weights only. KV cache, the OS and the 658MB vision projector come on top, so Unsloth says start at 32K context, not the full 1M
- You need a llama.cpp or Unsloth build that knows the new inkling2 architecture. Builds that ran the first Inkling may not load it
If you don't have roughly 200GB of RAM plus VRAM to spare, the model card lists hosted options: Together, Baseten, Nebius, Modal, Cloudflare Workers AI, Fireworks and NVIDIA.
Sources: huggingface.co/thinkingmachi… and unsloth.ai/docs/models/inkli…
#LocalAI #OpenWeights
1
50
llama.cpp hit v1.0.0 on 13 Nov. From now on, a breaking change to the public C API means a major-version bump, so pinning to 1.x should be safe.
Before you upgrade a setup that already works, check these four:
1. llama-server with tools or MCP on a non-loopback --host and no API key is now refused with an error. Set --api-key, or opt in with --allow-unauthenticated-tools.
2. Media URLs pointing at your own LAN (a NAS, say) are blocked unless you pass --media-allow-private-urls. The CORS proxy now only reaches public addresses, and there's no flag to undo that.
3. -dt / --defrag-thold is gone. Take it out of your launch scripts.
4. If you link libllama: 35 deprecated functions and llama_set_warmup are removed, llama_set_embd_state_src is renamed llama_set_ctx_state_src, and clip.h is folded into mtmd.
Sources: the v1.0.0 release notes and PRs #32113, #32115 and #30565 on GitHub (ggml-org/llama.cpp), plus the Hugging Face blog post announcing v1.0.0.
#llamacpp #LocalAI
1
20
Most Dec 1 Gemini posts stop at "2.5 Pro reroutes to 3.8 Flash". Three details in Google's own docs matter more if you run agents or pipelines.
1. Your temperature setting stops doing anything. Google's docs say temperature, top_p and top_k are ignored on the newer models, not rejected. A temperature 0 extraction job won't error, it just loses the setting. Google says future models will return a 400 instead.
2. gemini-2.5-flash lands on 3.5 Flash-Lite, whose default thinking level is minimal. Google's migration note says tool-using subagents should set medium or high to prevent premature tool termination.
3. Pro gets cheaper for one month, then not always. 2.5 Pro is $1.25 in and $10 out per 1M tokens (prompts up to 200k). 3.8 Flash is $0.75 and $3.75 until Dec 31, then $1.50 and $7.50.
Worked example, 180k tokens in, 2k out:
2.5 Pro today: $0.245
3.8 Flash in December: $0.1425
3.8 Flash from Jan 1: $0.285
Input-heavy calls like this end up costing more than they do now. Google also says 3.8 Flash uses more tokens on complex tasks by design, so output can grow too.
One hard failure as well: a request ending in a prefilled model turn now gets a 400.
What to do before Dec 1: set the model ID yourself, strip the sampling params, set thinking_level explicitly and rerun your evals. This is from the docs, I haven't run the reroute myself.
Google's deprecations page: aistudio.google.com/docs/dep…
#GeminiAPI #AIAgents
25
Both McLarens were thrown out of Las Vegas last year for worn skids. The limit they broke has changed for 2026.
2025 (Technical Regulations Article 3.5.9): plank 10mm thick when new, 9mm minimum after wear.
2026 (Section C, Article C3.6.1): still 10mm new, but the minimum is now 8mm.
The FIA stewards' Vegas 2025 decisions list the readings. Norris's lowest was 8.88mm. Piastri's was 8.74mm. Both under 9mm, both above 8mm.
That comparison is our own arithmetic, not an FIA statement. The stewards also say a re-measure came in lower still and give no figures, so treat it as rough.
How it's checked now: the skid comes off the car and gets measured with a micrometer around each of the three plank holes. The whole edge of every hole has to be inside the limit, so one thin spot is enough.
What it doesn't tell you: these are 2026 cars built to a different rulebook, so last year's wear says nothing about this year's. My read is that a bigger allowance mostly gets spent on running the car lower, which is exactly what the plank is there to police.
FIA 2026 Section C: fia.com/system/files/documen…
#F1 #LasVegasGP
39
Phone Pilot: free @bot and @grok template. Your bot drives iPhone and Android from a Windows PC, no Mac, several at once. Tested on Safari, YouTube, X and Mail. Write-up on my profile.
Drop what breaks in the replies, plus your phone model.
x.ai/bot/p4_Hc9ZKYElvEFm8I62…
#AIAgents
"Half of Ethereum blocks are now protected" is the wrong way to read the Phylax and Titan news. What protects an app is whether it installed an onchain guard.
Since 10 Nov, apps can register Solidity rules that Titan Builder checks while it builds blocks, and a transaction that breaks one is left out. Phylax puts Titan at roughly half of Ethereum blocks.
On its own, that just sends an attacker to another builder. So Phylax adds a guard: a protected function checks whether the current block carries a marker from an authorized builder, and reverts if it doesn't. Today that means Titan blocks. Phylax's Credible RPC routes transactions to participating builders. Send one another way and another builder may include it, where it reverts and still costs gas.
Three limits from Phylax's own docs:
1. Fail-open. A guard can be set to let calls through after a gap with no marked blocks. In that window other builders can include protected calls and run no rules.
2. The marker is the builder's word. The guard checks the marker, not a proof that the rules ran.
3. Rules only catch what they describe. Per Bankless, Denaria lost about $166K on Linea in April to an accounting bug its one rule (no bad debt) didn't cover.
Two more: a successful wallet simulation isn't proof the guard works, because zero gas price calls skip the marker check. And a rejected transaction isn't queued. Your wallet or relayer decides whether to retry.
If an app says it's protected, the useful questions are which functions are guarded, what the fail-open gap is, and which rules are live.
docs.phylax.systems/credible…
#Ethereum #DeFi
1
1
62
AWS Emulator matches real AWS far better at creating Lambda functions than at running them.
AWS released the free local emulator on 11 Nov and pitches it for coding agents that test against AWS. Each release ships a scores file. The public fidelity test suite labels every test control plane (create, describe, delete) or data plane (invoke, read, write, send).
Joining the two for release v1.0.322:
Lambda control plane: 138 of 159 pass
Lambda data plane, mostly invokes: 130 of 429
Tests that call a function URL: 0 of 30
For comparison on the data plane: SQS 217 of 230, S3 257 of 322, DynamoDB 371 of 465.
So a local run that deploys a function and invokes it tells you a lot less than one that writes to SQS or S3. AWS's docs say the same in general terms: control plane matches most closely, data plane varies by service.
To check your own app, list the AWS operations it calls, find them in the manifest file of the aws-emulator-evaluation-suite repo, then look up those test IDs in your release's scores file. If no test covers an operation, it isn't verified either way.
Caveats: this is a tally of AWS's published files, not an AWS figure. Single-service tests only, scores change with every release, and the emulator doesn't evaluate IAM policies the way AWS does.
docs.aws.amazon.com/emulator…
#AWS #AIagents

ALT Bar chart of AWS Emulator v1.0.322 fidelity tests passed by plane. Lambda: control plane 138 of 159, data plane 130 of 429. S3: 140 of 154 and 257 of 322. DynamoDB: 117 of 156 and 371 of 465. SQS: 69 of 74 and 217 of 230. Tally of the release scores file joined to the public test-suite labels, single-service tests only, not an AWS figure.
55



