- Published on
今日科技推文精选 - 2026-02-27
- Authors
- Name
- geeknotes
今日科技要闻：谷歌发布了更快速的图像生成模型 Nano Banana 2；与此同时，“下 7 个 Token 预测”技术的突破有望将推理速度提升五倍。虽然 AI 编程智能体正在根本性地改变编程工作流并极大地提高开发效率，但人力代价也已显现——杰克·多西（Jack Dorsey）旗下的公司爆发了大规模裁员，工程师面临的压力日益增加。此外，五角大楼正要求获得对 Anthropic AI 的无限制军事访问权限，这凸显了前沿模型、国家防务以及向全 AI 原生企业转型之间日益紧密的交集。
1. rseroter (Group Score: 155.8 | Individual: 43.0)
Cluster: 7 tweets | Engagement: 1170 (Avg: 163) | Type: Tech
RT @sundarpichai: Introducing Nano Banana 2, our best image model yet 🍌🍌
It uses Gemini’s understanding of the world and is powered by rea…
See 6 related tweets
- @CSProfKGD: RT @GoogleDeepMind: We’re launching Nano Banana 2, built on the latest Gemini Flash model. 🍌
It’s ...
- @GeminiApp: RT @Google: Introducing Nano Banana 2: Our best image generation and editing model yet. 🍌
Pro-level...
- @mark_k: Nano Banana 2 is coming TODAY on Thursday from @GoogleDeepMind 🍌
Technically the model is "Gemini 3...
- @testingcatalog: BREAKING 🚨: GOOGLE HAS ANNOUNCED NANO BANANA 2, A SOTA IMAGE GENERATION MODEL BASED ON THE LATEST GE...
- @googledevs: RT @googleaidevs: Nano Banana 2 (aka Gemini 3.1 Flash Image) is our SoTA model that offers image gen...
2. business (Group Score: 117.7 | Individual: 38.6)
Cluster: 5 tweets | Engagement: 218 (Avg: 81) | Type: Tech
Google is rolling out a new version of its popular AI image generation tool that’s meant to produce better visuals more quickly, six months after challenging OpenAI with the release of the original Nano Banana product https://t.co/etieccXneW
See 4 related tweets
- @testingcatalog: Nano Banana 2 is now available on Google AI Studio, along with a new "Image Search" tool!
"The mode...
- @Reuters: Google rolls out Nano Banana 2 after viral success of AI image generation tool https://t.co/bdyUciqF...
- @arstechnica: Google releases Nano Banana 2 AI image generator, promises Pro results with Flash speed https://t.co...
- @TechCrunch: Google launches Nano Banana 2 model with faster image generation https://t.co/2yxiVHTysi...
3. jerryjliu0 (Group Score: 95.6 | Individual: 58.1)
Cluster: 2 tweets | Engagement: 410 (Avg: 55) | Type: Tech
The Model Harness is Everything
We are already living in a world of incredible frontier models and incredible agent tools (Claude Code, OpenClaw). But the biggest barrier to getting value from AI is your own ability to context and workflow engineer the models. This is especially true the more horizontal the tool that you’re using.
If you’re using a very generic tool like ChatGPT and Claude Code, you need to spend a lot of work clearly articulating your requirements and specifications so that the agent can actually solve the task relative to your specifications. Today that looks like being extremely thoughtful about the tools that you select, and writing English very precisely in a https://t.co/GjvKCuoFDJ file to articulate the agent these requirements.
Some of the work around defining the business workflow is inherently time consuming. Think about any document SOP - simply writing the English can take hours to refine, iterate, and optimize. This is where more vertically focused agents come in; they handle the burden of equipping the agents with relevant prompts to solve a given workflow, so that you can just go in and use the application directly.
Another approach is to be specialized services that offer context to these agents. This is the space that we (@llama_index) are operating in. We are providing the infrastructure to parse the most complex documents into agent-ready context. For other companies it could be offering web data, sales data, documentation, or codebases as a service.
At a high-level any AI startup should provide context or workflows on top of these agents. We’re excited about building enduring tech even as the agent landscape evolves.
If you’re specifically excited to unlock the billions of context stored within your documents, come talk to us! https://t.co/Ht5jwxSrQB
See 1 related tweets
- @omarsar0: At this point, "agentic engineering" has allowed me to build the best AI harness I could possibly ge...
4. garrytan (Group Score: 74.6 | Individual: 41.8)
Cluster: 2 tweets | Engagement: 5952 (Avg: 344) | Type: Tech
RT @karpathy: It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the…
See 1 related tweets
- @aakashgupta: The buried lede in this post is the October-to-December timeline.
In October 2025, Karpathy publicl...
5. business (Group Score: 73.1 | Individual: 26.4)
Cluster: 4 tweets | Engagement: 114 (Avg: 81) | Type: Tech
AI coding agents promised to make software development easier. Instead they’re pushing engineers to work faster and longer https://t.co/yLTb7v21xC
See 3 related tweets
- @WIRED: Silicon Valley built AI coding agents that can handle most of the grunt work. Now, the most valuable...
- @javarevisited: We built AI to automate coding. Now developers spend more time reviewing AI code....
- @WIRED: RT @ZeffMax: Silicon Valley has long prized its high agency individuals. But now that AI coding agen...
6. Cointelegraph (Group Score: 69.7 | Individual: 16.9)
Cluster: 5 tweets | Engagement: 550 (Avg: 562) | Type: Tech
🇺🇸 JUST IN: Pentagon issues “best and final” offer to Anthropic, demanding unrestricted military use of its AI ahead of a Friday deadline. https://t.co/fwhHvbVj6c
See 4 related tweets
- @Reuters: 🔊 ‘The Pentagon and Pete Hegseth really underscored that they still want to use Anthropic's AI, but ...
- @Reuters: 🔊 ‘This is coming at a time when the Pentagon is writing the initial foundational contracts of how A...
- @business: Anthropic's high-stakes conflict with the Pentagon gets to the heart of just how far AI can shape th...
- @Miles_Brundage: RT @JenGriffinFNC: Strange that the Pentagon/Sec Hegseth picks this fight with Anthropic, the AI com...
7. business (Group Score: 49.5 | Individual: 22.2)
Cluster: 3 tweets | Engagement: 52 (Avg: 81) | Type: Tech
Over the past year or so, Wall Street has gone through waves of AI-related selloffs. This week’s market dip may have been the first partly caused by a self-published work of fiction. Read more on how the reaction is the latest indication yet that traders are struggling to wrap their head around the trajectory for AI: https://t.co/uMXE7IX1yL
📷️: Michael Nagle/Bloomberg
See 2 related tweets
- @business: Wall Street has been dumping software stocks over artificial intelligence disruption fears, but non-...
- @FT: Wall Street turns to complex trades to dodge AI ‘implosions’ https://t.co/vVPdPFTuhV...
8. BrianRoemmele (Group Score: 49.0 | Individual: 33.2)
Cluster: 3 tweets | Engagement: 141 (Avg: 526) | Type: Tech
A 5x AI Speed Up With Not Next Token Prediction But NEXT 7 TOKEN PREDICTION!
Next-Token Prediction Just Got Retired: And I’m Already Running the Future in My Lab Right Now
I’ve been saying it for years: the autoregressive bottleneck is the single biggest drag holding back real-time, production-scale AI. One token at a time? That’s over.
In a new paper researchers took pretrained models, specifically Llama-3.1-8B-MagpieAlign-SFT-v0.1 and Qwen3-4B-Instruct and turned them into native multi-token predictors using nothing more than a simple online self-distillation objective.
No extra draft models.
No speculative decoding scaffolding.
No verifier.
No new architecture.
Just the exact same weights and implementation as the original checkpoint… now spitting out 2–7 tokens (sometimes more) in a single forward pass.
They call the inference trick Confidence-Adaptive Decoding (ConfAdapt). The model dynamically decides how many tokens it’s confident enough to commit to. High-confidence spans fly out in chunks. Tricky spots fall back to single-token precision. It’s like the model is self-regulating its own speed vs. quality trade-off in real time.
On GSM8K (grade-school math, the classic reasoning benchmark):
- Llama-3.1-8B variant: >3× faster decoding with <3% accuracy drop by (τ=90% confidence threshold).
- Up to 5×* acceleration if you’re willing to accept a bit more trade-off.
- Average chunk size ~3–6 tokens per forward pass in practice.
And the quality holds across instruction following, open-ended generation, and other reasoning suites. This isn’t “fast but dumber.” It’s fast and almost indistinguishable.
Figure 1 in the paper shows a beautiful GSM8K solution with colored blocks of 1–7 tokens generated at once. Average chunk size: 3.04. Pure poetry.
This Is a Genuine Paradigm Shift
Speculative decoding? Cool, but you need a whole extra model and fragile pipelines.
Medusa / Lookahead? More scaffolding.
This? You literally distill the model against its own frozen teacher copy in an on-policy RL-style loop. The student learns to predict spans that the teacher would have produced anyway. Then at inference… it just works. Drop-in replacement.
The authors nailed it: “Future architectures will be optimized for sequence compression and throughput, not token latency.”
I’ve been screaming this exact sentence since 2023. Today it’s not theory, it’s downloadable checkpoints.
I’m Testing It RIGHT NOW (Feb 26, 2026, Live From the Lab)
As soon as the checkpoints hit Hugging Face (https://t.co/HwRc7XgeXY), I spun them up.
First run: Llama-3.1-8B-MTP variant on a long-form reasoning chain I use daily.
Wall-clock speedup: 3.4× on my A100 setup.
Coherence? Identical to baseline for 95%+ of outputs.
I threw it at a 4,000-token agent workflow that normally takes 18 seconds, now under 6 seconds.
I’m already wiring it into The Zero-Human Company.
What This Means for All of Us
- Inference costs just got slashed.
- Real-time voice agents that actually feel instant? Finally.
- Longer reasoning chains without blowing your budget? Trivial.
- The entire “optimize the decoder” cottage industry just got disrupted overnight.
We’re not waiting for 100T-parameter monsters anymore. We’re making the models we already have radically more efficient at the architecture level.
Next-token prediction didn’t die today. It was mercy-killed, cleanly, elegantly, and with reproducible code.
The throughput wars just began.
And I’m all in.
Paper: https://t.co/8X1PiBedhm
Checkpoints: https://t.co/HwRc7XgeXY
Code: https://t.co/QmbS9r4PUF
See 2 related tweets
- @rohanpaul_ai: RT @rohanpaul_ai: Multi-token prediction via self-distillation delivers 3x inference speedups in mod...
- @BrianRoemmele: RT @BrianRoemmele: A 5x AI Speed Up With Not Next Token Prediction But NEXT 7 TOKEN PREDICTION!
Nex...
9. BrianRoemmele (Group Score: 43.3 | Individual: 22.1)
Cluster: 2 tweets | Engagement: 70 (Avg: 526) | Type: Tech
The Zero-Human Company @ Home Potential Is Massive!
10 teraFLOPS minimum!
The Zero-Human Company endeavors to captures unused CPU and GPU idle clock cycles at enterprise scale through secure local execution.
Here is a near term target of one million participating nodes each delivering an average of ten teraFLOPS of sustained optimized AI compute.
Aggregate compute calculation
Nodes 1 000 000
TFLOPS per node 10
Total TFLOPS 1 000 000 * 10 = 10 000 000 TFLOPS
Conversion note 1 exaFLOPS = 1 000 000 TFLOPS
Total exaFLOPS 10 000 000 / 1 000 000 = 10 exaFLOPS
This aggregate matches or exceeds the combined public benchmark output of the TOP500 supercomputers which stood at approximately fourteen exaFLOPS in mid 2025.
See 1 related tweets
- @BrianRoemmele: The Academic paper on The Zero-Human Company @ Home.
We are aiming for 10 teraFLOPS minimum with: Z...
10. dbreunig (Group Score: 42.2 | Individual: 24.5)
Cluster: 2 tweets | Engagement: 28 (Avg: 126) | Type: Tech
On March 18th, we're hosting another Bay Area DSPy Meetup featuring in-production case studies involving GEPA, tool use, and LLM judges from Dropbox and Shopify. (And we'll talk RLMs, too.) Join us! https://t.co/SnmobNhXup
See 1 related tweets
- @lateinteraction: RT @dbreunig: On March 18th, we're hosting another Bay Area DSPy Meetup featuring in-production case...
11. K8sArchitect (Group Score: 41.8 | Individual: 25.4)
Cluster: 2 tweets | Engagement: 35 (Avg: 12) | Type: Tech
sk8r is a modern open source Kubernetes dashboard built with Svelte and TypeScript that provides a user-friendly interface to visualize and manage cluster resources, stream pod logs in real time, and monitor metrics through Prometheus integration
➜ https://t.co/GbifN8NFUt https://t.co/REMTQ8Eags
See 1 related tweets
- @learnk8s: RT @K8sArchitect: sk8r is a modern open source Kubernetes dashboard built with Svelte and TypeScript...
12. minchoi (Group Score: 40.8 | Individual: 34.6)
Cluster: 2 tweets | Engagement: 1146 (Avg: 166) | Type: Tech
This AI ad is wild 😂 https://t.co/Oy2Sencny8
See 1 related tweets
- @EHuanglu: AI ad is next level https://t.co/TkcixdEfWV...
13. AlexFinn (Group Score: 40.8 | Individual: 40.8)
Cluster: 1 tweets | Engagement: 2728 (Avg: 1037) | Type: Tech
Jack Dorsey just laid off half of his company in a single tweet. 4,000 people gone
Not because business is down
But because AI made them unnecessary
If you aren’t AI native, you have become expendable to execs.
You need to learn these skills now:
- How to build software in Claude Code
- How to automate in OpenClaw
- How to create artifacts in Claude Cowork
- How to orchestrate multiple agents in Codex
- How to use ChatGPT as a copilot for everything you do
These aren’t optional skills anymore. They’re mandatory.
And the time you have left to learn them has quickly disappeared.
14. mattturck (Group Score: 39.5 | Individual: 32.3)
Cluster: 2 tweets | Engagement: 25 (Avg: 243) | Type: Tech
Can AI be correct 100% of the time? Verification as the missing layer for reasoning superintelligence - my conversation with @CarinaLHong, the incredibly impressive CEO of @axiommathai.
00:00 Intro
01:25 Why the World Needs an AI Mathematician
02:57 Scoring 12/12 on the World's Hardest Math Test (Putnam)
04:05 The First AI to Solve Open Research Conjectures
06:59 Does AI Solve Math in "Alien" Ways? (The Move 37 Effect)
08:59 "Lean": The Programming Language of Proofs Explained
10:51 How Axiom's Approach Differs from DeepMind & OpenAI
16:06 Formal vs. Informal Reasoning (And Auto-Formalization)
17:37 The AI "Reward Hacking" Problem
20:18 Building an AI That is 100% Correct, 100% of the Time
23:23 Beyond Math: Verified Code & Hardware Verification
25:12 The Brutal Reality of Competitive Math Olympiads
29:30 From Neuroscience to Stanford Law to Dropout Founder
33:57 How Axiom Actually Works Under the Hood (The Architecture)
37:51 The Secret to Generating Perfect Synthetic Data
40:14 Tokens, Proof Length, and Inference Cost
42:58 The "Everest" of Mathematics: Scaling Reasoning Trees
46:32 Can an AI Win a Fields Medal?
47:25 "Math Renaissance": What Changes if This Works
55:47 How Mathematicians React to AI (And Why Proof Certificates Matter)
57:30 Becoming a CEO: Dropping Ego and Building Culture
1:00:42 Recruiting World-Class Talent & Building the Axiom "Tribe"
See 1 related tweets
- @minchoi: RT @minchoi: Ok AI math is insane...
An AI just solved 6 of 10 unsolved research math problems... f...
15. nummanali (Group Score: 38.9 | Individual: 38.9)
Cluster: 1 tweets | Engagement: 224 (Avg: 47) | Type: Tech
I think I've found the best ADE for coding agents
- Native Ghostty Lib
- GPU Accelerated
- Notifications support
- Keyboard shortcuts
- Supports all agents
- Browser built in
- MacOS Native
- Zero Worktrees
Website clean, docs are on point and OSS
https://t.co/cR5oUYzGcG https://t.co/penRZbCRkS
16. aakashgupta (Group Score: 38.5 | Individual: 38.5)
Cluster: 1 tweets | Engagement: 791 (Avg: 466) | Type: Tech
A company with 12.2 billion in gross profit. Stock ripped 20% after hours. The market added roughly 1.5 million in enterprise value created per eliminated role.
Block is the canary in the coal mine. And they're not alone.
ASML cut 1,700 jobs last month while reporting record orders and said they were "choosing to make these changes at a moment of strength." Salesforce cut 5,000 after AI agents started handling 50% of customer interactions. Amazon cut 16,000 in January on top of 14,000 in October. Every one of these companies was growing when they did it.
Dorsey said the quiet part out loud: intelligence tools paired with smaller teams have already changed what it means to run a company. He chose one massive cut over repeated rounds because, his words, gradual cuts destroy morale and trust. The restructuring charges are $450-500 million. At the operating income Block is guiding, that pays for itself in two quarters. After that, pure margin expansion. That's why Wall Street rewarded it instantly.
Here's what's coming. Goldman estimates AI is already responsible for 5,000 to 10,000 net monthly job losses in exposed U.S. industries. Citigroup is planning 20,000 cuts. Dow just slashed 4,500. 40% of employers surveyed say they expect to reduce headcount because of AI. 30,700 tech jobs gone in the first six weeks of 2026 alone.
Block went from 10,000 to 6,000 while growing revenue and raising guidance. Every CEO running a company with more than a few thousand employees is doing this math tonight. The canary just stopped singing.
17. minchoi (Group Score: 38.2 | Individual: 31.0)
Cluster: 2 tweets | Engagement: 404 (Avg: 166) | Type: Tech
Programming will never be the same... 🤯
Karpathy gave an agent one sentence. 30 min later...
SSH set up, vLLM running, endpoints live, web UI built. He didn't touch anything.
This is the new normal now.
See 1 related tweets
- @minchoi: RT @minchoi: Programming will never be the same... 🤯
Karpathy gave an agent one sentence. 30 min la...
18. kimmonismus (Group Score: 37.5 | Individual: 22.3)
Cluster: 2 tweets | Engagement: 306 (Avg: 422) | Type: Tech
Amazon’s $50B AGI or IPO Gambit on OpenAI (outplaying microsoft?)
Amazon is negotiating a potential 15B upfront and another $35B tied to either an IPO or achieving artificial general intelligence (AGI).
OpenAI’s compute costs are projected to hit $665B over five years, pushing it toward public markets and deeper cloud partnerships, including expanded use of Amazon’s Trainium chips and custom AI models for Alexa. If OpenAI reaches AGI, Microsoft’s exclusive Azure hosting rights could loosen, unlocking huge strategic upside for Amazon.
See 1 related tweets
- @jukan05: The information: OpenAI is in talks to significantly expand its previously announced $38 billion clo...
19. rohanpaul_ai (Group Score: 37.0 | Individual: 29.7)
Cluster: 2 tweets | Engagement: 79 (Avg: 110) | Type: Tech
This new paper introduces a benchmark proving that human-written skill guides dramatically improve AI agent performance.
Standard LLMs often struggle with specialized industry tasks because they lack the exact step-by-step knowledge required to complete them.
Instead of permanently altering the model through expensive training, developers are now giving these agents plain text files called skills that explain exactly how to handle specific workflows.
Researchers built a massive test containing 84 complex tasks to see if providing these external skill documents actually helps 7 different popular AI models succeed.
They discovered that handing the AI a human-curated skill file boosts its passing score by 16%, but asking the AI to write its own guide actually hurts its performance.
This shows that while AI models are powerful, they still completely depend on human experts to map out the exact procedures for highly specialized jobs.
Paper Link – arxiv. org/abs/2602.12670
Paper Title: "SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks"
See 1 related tweets
- @rohanpaul_ai: RT @rohanpaul_ai: This new paper introduces a benchmark proving that human-written skill guides dram...
20. paulg (Group Score: 36.6 | Individual: 27.3)
Cluster: 2 tweets | Engagement: 1056 (Avg: 894) | Type: Tech
For the foreseeable future, everything about starting a startup, both good and bad, will be accentuated. It will be even harder to figure out what to do, but the founders who get it right will be able to create amazing things even faster than they could before.
See 1 related tweets
- @ycombinator: RT @paulg: For the foreseeable future, everything about starting a startup, both good and bad, will ...