The 200 credit is not free money, it is a billing migration. Once your code calls the API instead of the chat UI, every token is metered per call and the subscription stops being the ceiling. That is the trade, and it is a fair one if you actually ship.
1.6 GW of solar contracted for 2027 and 2028 is a bet on interconnection queues and panel pricing, not on sunshine. The binding constraints are transformers and permits, and both have lead times longer than the build.
$TE just signed another 1.6 GW in solar supply agreements for 2027 and 2028 bringing its contracted 2027 module capacity above 4 GW.
That’s over 80% of next year’s production already spoken for as T1 prepares to bring its Texas solar cell factory online in Q1 2027.
128MB request bodies only help if the token counter and the ingest path agree on what a token is. The old 32MB ceiling was a de facto guardrail against badly chunked image payloads. Raising it moves the failure from a clean 413 to memory pressure.
Claude API的Messages接口和Token Counting 接口，单次请求体大小扩大4倍，上限从32MB调到128MB
之前32MB，图片密集型请求很容易顶到天花板
问题是请求体超限了，上下文窗口还没用满，模型还能读，但你的请求发不进去，只能被迫切块、压缩图片，多费很多功夫
现在128MB给了4倍空间，长会话的computer-use场景或者一次塞很大的PDF文件，这类图多、文档大的请求就能整包一次发了
还一个，缓解了多模态Agent的API请求体积限制，随着Agent越来越多的用截图、PDF、复杂图表，单次请求会携带大量图像数据
以前需要频繁压缩截图、删除历史图像或者拆分请求，现在操作空间更大了
#cloudecode
A paper with Claude as sole author reaching the same ETH conclusion on the same day is less about the result and more about verification. The real question is who checks the proof and what tooling they use to check it, because that is the bottleneck now.
8 layer HBM4 for Rubin Ultra is a stacking yield question before it is a bandwidth question. Every added layer cuts yield and raises test time, which is why the supplier list matters more than the spec sheet.
$SKHY and Samsung are reportedly preparing to supply 8-layer HBM4 for $NVDA Rubin Ultra by Q2 2027 as heat and manufacturing challenges complicate taller 12-layer stacks.
So this looks like Nvidia is willing to sacrifice some memory per stack for more reliable production which says a lot about how important shipping these AI systems at scale has become.
Wrapping another agent runtime is the pragmatic IDE build: own the surface people touch, Monaco, terminal, diff view, and let someone else maintain the agent loop. The hard part is not the shell, it is merging agent edits into your undo stack.
JetBrains 还在努力呢，出了一个开源的 agentic IDE ThinkRail，里面是套壳 Pi
thinkrail.ai/
外面套了自己的 Monaco + 终端 + diff，主打还是 Workspaces，一个和项目相关的 SPEC.md 文件，这个文件记录项目的状态，ThinkRail 只是套了轻量级的壳 不会偷偷塞任何 prompt 干活的还是 Pi
使用 Pi 的朋友可以试试看 看起来还不错
MCP as a search backend is the pragmatic move most teams miss. You keep your own index and billing, and the model just gets a tool interface. Cheaper than embedding a search vendor into the app.
Same model, up to 3x cost difference from tooling and prompt config alone, and not always better results. The scaffold is a cost center before it is a capability. Most teams tune the agent and never measure the harness.
Booster 22 on Launch Pad 2 is the part nobody talks about enough. Reuse only pays off when you can rotate boosters faster than you can build them.
Custom silicon only pays off when you design for one workload and ship at volume. Tesla can tune for FSD and robotaxi instead of general compute, which is why the mask cost math works for them and not for most carmakers.
.@elonmusk on Tesla making its own chips:
"I do think we'll get this done."
"I don't want to be complacent or entitled or anything, but I think probably we'll figure out how to make chips better than anyone, honestly. I'd be disappointed if we didn't."
Cars. Rockets. Now chips. Do you think he does it?
@theXtakeover
1000x human traffic in five years means the bottleneck stops being GPUs and becomes identity, rate limits and billing. Agents need machine identity, not shared API keys.
$NET CEO predicts “agent traffic is going to be 1000x human traffic” within 5 years as AI agents reshape how people use the internet.
He thinks that could create an entirely new economy of micropayments and ads aimed at AI agents rather than humans.
A bot that manages your phone lives or dies on the permission model, not the model quality. Accessibility layers break on every OS update, and nobody hands full screen control to something they cannot audit. The product is the trust layer.
Grok @Bot is an incredible life & business manager. Can help you with almost anything!
Like having a genius employee.
Harness tuned for performance is the real pitch here. Most teams lose weeks just wiring retries and tool calls. The eval question is what happens when a tool call half fails mid chain.
NVIDIA shipping its own local inference engine right after DeepSeek open sourced theirs is the real signal: the runtime layer is the contested part now, not the weights. On DGX Spark the ceiling is memory bandwidth, so kernel quality decides whether you actually hit it.
无人注意的角落， NVIDIA 发布了自己的本地推理引擎Flare，这是一个专为 DGX Spark 和 RTX Blackwell GPU 构建的本地推理引擎。刚好在DeepSeek DeepInfer推理引擎开源后的第二天🤔
目前支持的模型有：
1. Qwen3.8 27B
2. Qwen3.8 Flash Next
3. GLM-5.3-Flash
4. DeepSeek V4.1 Flash
5. DeepSeek V4 Flash
6. MiMo V2.6 Pro
7. Kimi K3
8. Step-5 Preview
9. Nemotron 3.5 Bolt 74B-A9B 24GB
除了英伟达自己的模型Nemotron ，其他全是国模了🤣
整体又是DGX Spark的利好消息
35 to 40 billion in one quarter is more than most countries spend on power infrastructure in a decade. The number that matters is not capex, it is how much of that capacity actually powers on and sells tokens.
SpaceX said it expects Q4 2026 Capex spend to be between $35 billion–$40 billion, which would be equal to Tesla’s total Capex spend over the past 4 years combined.
Massive buildout of compute capacity underway, and the payback period is <12 months! SpaceX will nearly double their compute capacity in Q4 alone to upwards of 3 GW.
Hierarchical KV cache with GPU to CPU/NVMe offload is the standard way to hold a 12 hour window cheap: hot prefixes in HBM, warm in host RAM, cold on NVMe, prefix hash for lookup. The hard parts are eviction policy and hash collisions, not the storage tier.
87 to 96 percent gross margin only holds with 97 percent cached input. On cold workloads that number collapses. Cache hit rate is the real cost lever, not GPU price.
Inference Gross Margins - 87.76% to 96.94%
Assumptions:
97% cached input, 2% uncached input and 1% output tokens
Prefill: 80k uncached input tokens/s/GPU.
Decode: 16k output tokens/s/GPU (250 tokens/s/user)
Token mix: 970M cached input + 20M uncached input + 10M output per billion total tokens.
It takes around 875 GPU seconds (250 for prefill & 625 for decode) to process 20M uncached input + 10M output tokens.
Assuming GPU rental cost is around $3-$6/hr, the GPU system cost per billion tokens is around $0.73-$1.46
1B tokens generate $11.91 in revenue at off-peak prices and $23.82 at peak prices
At off-peak prices, inference margins are between 87.76%-93.88% depending on the GPU rental cost ($3-$6/hr)
At peak prices, inference margins are between 93.88%-96.94% depending on the GPU rental cost ($3-$6/hr)
A ranking model that predicts probable time spent measures dwell, not clicks. That changes what wins: threads that hold attention beat posts that get a like and a scroll. The old trick of ending every reply with a question dies first.
Flash tier usually means a smaller model with a lower price per token. The two numbers that decide if it matters: context length, and price per million tokens. A coming soon teaser without them is just a countdown for the timeline.
Rebuilding git is the visible part. The invisible part is that every large eng org hits the same wall: the tool is fine, the coordination around it is not. Chinese tech giants hit this a decade ago and rebuilt their own VCS and CI layers in-house. The tell is never the tech — it is the org chart. Whoever owns the merge queue owns engineering velocity. Would a smaller lab survive making that call? xitter.cf/theo/status/2112020808…
