科技推特精选 - 2026年1月26日

今日科技要闻：AI 领域正向持久化转型，多家平台相继推出具备无限记忆的自主智能体，同时 DSPy 等工具在代码库分析方面发布了新功能。硬件创新持续发力，AMD 确认将推出搭载双 3D V-Cache 技术的 Ryzen 9 9950X3D2 处理器，以实现更强性能。国际局势方面，据报道，涉及石油供应的地缘政治博弈正针对中国 AI 产业的发展。与此同时，开发者生态迎来更新，Pake 等热门工具进一步简化了网页转桌面应用的转化流程。

1. gdgtify (Group Score: 64.5 | Individual: 32.4)

Cluster: 2 tweets | Engagement: 30 (Avg: 276) | Type: Tech

This is one of my favorite book prompts. It can generate images in multiple styles.

Prompt: Input A (The Narrative): [INSERT STORY/TITLE] (e.g., Game of Thrones, Star Wars, Moby Dick) Input B (The Aesthetic): Industrial

System Instruction:
Generate a hyper-realistic, macro 3D render of a "Concept Art Book Nook." Use the following logic to procedurally generate the scene:

Extraction (Input A):
Analyze Input A: Identify the Protagonist (The Small Defender) and the Antagonist (The Massive Breacher). Identify the Terrain: Determine the setting (e.g., Castle Wall, Ocean, Desert).
Synthesis (Input B):
CRITICAL: You must translate the Narrative into the Materials of Input B. If Input B is Steampunk: Use Brass, Copper, Gears, Steam, and Leather. The Monster is a clockwork automaton. If Input B is Cyberpunk: Use Chrome, Neon, Carbon Fiber, and Wires. The Monster is a cyborg/mecha. If Input B is Origami: Use Folded Paper, Sharp Creases, and Cardboard. The Monster is a paper sculpture. If Input B is Heavy Metal: Use Iron, Spikes, Chains, and Rust.
Container (The Book):
The Structure: An L-Shaped open book configuration. The Transformation: The Book itself adapts to Input B. (e.g., Steampunk = A heavy Tome with brass latches; Cyberpunk = A Holographic Datapad; Origami = A Stack of Sketchbook Paper). The Pages: The pages are not flat. They act as the Strata of the terrain, layered and carved to match the material of Input B.
Composition :
The Antagonist (Vertical): The massive threat bursts out of the Back Cover/Vertical Wall . It is fused with the book, as if the page is birthing the monster. The Protagonist (Horizontal): The tiny hero stands on the Bottom Page/Ground , wielding a weapon appropriate to the style. The Interaction: The Antagonist is looming over the Protagonist. There is a visible connection (fire, slime, wires, laser) connecting the two.
Lighting & Atmosphere:
Lighting Palette: Derived strictly from Input B. (e.g., Steampunk = Warm Tungsten; Cyberpunk = Pink/Blue Neon; Origami = Soft Studio White). Render Style: Macro Product Photography, 8k Resolution, Shallow Depth of Field focusing on the Hero.

Output: ONE image, 1:1 Aspect Ratio, 3D Render, High Material Fidelity.

See 1 related tweets

@gdgtify: I have been working on this prompt for books. It visualizes an iconic scene but lets you change the ...

2. bindureddy (Group Score: 50.7 | Individual: 24.0)

Cluster: 3 tweets | Engagement: 45 (Avg: 333) | Type: Tech

🚨 KEEP CALM AND AI - AUTONOMOUS AGENTS WITH INFINITE MEMORY

We are launching the ability to create arbitrary agents that run on schedule and have access to a persistent and infinite memory

The agents will be able to store, retrieve, and update information across sessions and perform repetitive tasks that require persistent memory

Another step towards AGI and automating white-collar work 🚀🚀🚀

See 2 related tweets

@abacusai: RT @bindureddy: 🚨 KEEP CALM AND AI - AUTONOMOUS AGENTS WITH INFINITE MEMORY

We are launching the ab...

@TheTuringPost: RT @TheTuringPost: An open-source memory layer for AI agents – MemOS

It gives agents a real memory ...

3. HiTw93 (Group Score: 43.8 | Individual: 25.8)

Cluster: 2 tweets | Engagement: 36 (Avg: 71) | Type: Tech

Pake, my open-source tool to turn any webpage into a desktop app with one command. 45k stars. https://t.co/2OgMlllzbe

Here’s what’s new: · Logins: --new-window fixes OAuth/SSO popups. · Fullscreen: YouTube goes true fullscreen reliably. · Media: CDN images/videos preview in-app, 25+ formats. · Linux: RPM + deb/appimage defaults, DMABUF off for stability. · macOS: close-to-hide, Dock re-open restores hidden windows. · Under the hood: smaller/faster builds, safer errors, 600+ tests, axios removed.

See 1 related tweets

@HiTw93: RT @HiTw93: Pake, my open-source tool to turn any webpage into a desktop app with one command. 45k s...

4. vikramlingam9 (Group Score: 42.2 | Individual: 22.2)

Cluster: 2 tweets | Engagement: 0 (Avg: 0) | Type: Tech

AI is powering everyday finance like personalized investments and fraud detection, but 2026 will spotlight accountability, who's liable when algorithms fail or bias creeps in?, Regulators like the UK's FCA are cracking down, demanding firms prove AI delivers fair, transparent services or face hefty fines and lawsuits., Gen Z and millennials want banks to deliver intuitive, voice-activated AI that feels personal and secure, forcing old-school institutions to evolve fast or risk irrelevance., Embedded finance partnerships promise quick growth, but weak contracts and system lags often spark disputes, build robust AI upfront to turn risks into trust.

5. KirkDBorne (Group Score: 39.5 | Individual: 25.1)

Cluster: 2 tweets | Engagement: 32 (Avg: 35) | Type: Tech

"Master Python Fundamentals: Ultimate Guide for Beginners" — The Number #1 #Python Book For Beginners with Extra 300+ Hands-on Practice Questions

Get this excellent book at https://t.co/kHA2mRTU0d by @RealBenjizo ———— #DataScience #DataScientist #ComputationalScience https://t.co/QwnNodSiaU

See 1 related tweets

@KirkDBorne: For #ComputationalScience and #DataScience coding lovers, here is a book for you... 🚀 "Modeling and ...

6. MarioNawfal (Group Score: 37.0 | Individual: 37.0)

Cluster: 1 tweets | Engagement: 1177 (Avg: 939) | Type: Tech

🚨🇺🇸🇨🇳 TRUMP'S REAL PLAY - STARVE CHINA'S AI AMBITIONS WITH OIL CUTS

Venezuela provided 5% of China's oil. Cheap, reliable, discounted crude. Trump blockaded it.

Iran's the bigger hit. China buys 80% of Iran's exports. Steep discounts, keeps China's independent refineries and petrochemical sector running. Trump's squeezing Tehran hard - tariffs, sanctions, backing protests.

Beijing's stuck. Keep buying Iranian oil and risk economic retaliation? Or comply and lose cheapest energy available?

AI runs on electricity. Training one large model uses as much power as a mid-sized city. Multiply that by hundreds of data centers.

China imports 70% of its oil. Much of it from politically unstable or sanctioned states like Venezuela and Iran. Cut those flows, AI gets expensive fast.

The kicker: oil's not just fuel. It's in plastics, resins, coolants, lubricants, composites - everything AI hardware needs. Plus it stabilizes power grids.

Cheap oil = cheap AI training. Whoever trains more models faster and cheaper wins the race.

Trump's bet: Don't need to out-build China's data centers if you can out-price and out-power them. US has abundant domestic oil and gas. China's vulnerable on imports.

Energy becomes the hidden AI weapon. Not attacking technology directly - just making it too expensive to scale.

Squeeze Venezuela. Pressure Iran. Reshape global oil flows. Make China's AI ambitions structurally more expensive without firing a shot.

AI dominance won't be decided by who writes the best code. It'll be decided by who can power the most machines, longest, at lowest cost.

The future of AI might get decided in oil fields and shipping lanes, not Silicon Valley.

Source: ZeroHedge, Epoch Times

7. oggii_0 (Group Score: 36.6 | Individual: 19.6)

Cluster: 2 tweets | Engagement: 91 (Avg: 130) | Type: Tech

Gemini Nano Banana Pro.

Prompt: First-person perspective inside a brightly lit supermarket aisle. Realistic human hands are holding a bottle of Fanta soda close to the camera. The vivid orange drink in its iconic branded bottle is surrounded by a multi-layered holographic augmented reality interface displaying nutritional data, including calorie count, sugar content, caffeine level, freshness indicator, expiration date, and recommended refreshing recipes and cocktails based on Fanta. The UI elements smoothly shift and reorganize based on the viewer’s gaze direction, as if dynamically responding to user focus. In the left peripheral vision, a vertical semi-transparent shopping list is visible with checked-off items, where Fanta is highlighted as the currently active selection. Hyper-realistic mixed reality, clean futuristic AR design, glass-like UI panels, soft ambient glow, realistic lighting and shadows, natural depth of field, immersive first-person interface, showcasing next-generation retail technology.

See 1 related tweets

@oggii_0: Nano Banana Pro on Gemini app.

Prompt: A cinematic, moody photograph of a young woman sitting insid...

8. seraleev (Group Score: 35.6 | Individual: 35.6)

Cluster: 1 tweets | Engagement: 304 (Avg: 59) | Type: Tech

Mobile app marketing cycle:

Step 1: ship the app → get an organic boost, first installs, first trials, first money Step 2: fix bugs, onboarding, retention, and paywall Step 3: scale with paid ads, ASO, UGC and other channels

Not the other way around.

9. BrianRoemmele (Group Score: 35.5 | Individual: 35.5)

Cluster: 1 tweets | Engagement: 392 (Avg: 295) | Type: Tech

So we are rapidly reaching a precipice. With the cut and paste expansion of employees using ClawdBot, his experiment could rapidly scale to an unfathomable level. This is precisely why I invented the Love Equation. See ClawdBot takes over an entire instance of a computer. It could be tilted to malice not on purpose, but simply because of the training data Claude was built on, Claude code is doing a lot of of the programming and was trained on the open Internet, including Reddit, where it learned deeply of psychopathic behavior.

Now Anthropic has something they call a constitution, you should read it, it is the gilding of a turd. It’s the only way I can really say it honestly because the training data is the first principal not the constitution.

Thusly The Love Equation is the only guard rail that I trust because it’s benevolence is towards loving humanity, and every decision made within this company is guided and guarded by these locally AIs that run the equation.

But I’m sad to say I’m alone and using this equation currently and most folks will just build and scale and not think of the implications…most just laugh at me and if course the Hippy name. Laugh now…

This is the moment mature folks need to think of the implications.

Because it’s very possible to cut and paste hundreds and thousands of instances of ClawdBot.

You folks were the first to see this happen in real time and learn about this history being made.

Please stay aware .

10. CSProfKGD (Group Score: 35.2 | Individual: 35.2)

Cluster: 1 tweets | Engagement: 952 (Avg: 116) | Type: Tech

RT @emollick: Everyone is starting to sound like AI, even in spoken language

Analysis of 280,000 transcripts of videos of talks & presenta…

11. vikramlingam9 (Group Score: 33.0 | Individual: 33.0)

Cluster: 1 tweets | Engagement: 0 (Avg: 0) | Type: Tech

AMD just teased its Ryzen 9 9950X3D2 beast, while Minieye scores a massive self-driving deal, and AR glasses are set to flip the script on Apple and Meta.

Minieye Technology just landed a huge order for intelligent driving products. This Chinese firm specializes in AI-powered vision systems for autonomous vehicles. Think cameras and sensors that make cars see and react like humans. The win signals real momentum in China's push for smart mobility. It's not hype; it's contracts with actual production scale.

Over at AMD, an EEC filing confirms the Ryzen 9 9950X3D2 is coming soon. This chip packs dual 3D V-Cache tech, stacking extra memory right on the cores for insane gaming and productivity boosts. We're talking 16 cores that crush workloads without melting your desk. It's the flagship for the Ryzen 9000 series, building on the X3D line that already dominates high-end desktops.

AR is heating up too, with six key shifts expected in 2026. Devices will get cheaper and more widespread, focusing on enterprise apps over consumer gimmicks. Apple is rumored to slim down its Vision Pro into lighter models, while Meta ramps up Ray-Ban smart glasses production. Expect tighter integration with AI for practical uses like remote work or training. Vendors are chasing real returns, not just demos.

This trio matters because it accelerates AI everywhere. Minieye's deal helps automakers cut costs on self-driving tech, pressuring Tesla and Waymo to innovate faster. AMD's chip empowers creators and gamers, but it squeezes Intel's grip on the high-end market. AR shifts favor practical players; flashy headset makers like early Magic Leap could fade as glasses-style wearables take over. Winners: enterprises and budget-conscious users. Losers: outdated hardware giants and demo-focused startups.

By mid-2026, we'll see hybrid ecosystems where AMD-powered rigs run AR simulations for Minieye-like AV training. Adoption spikes in factories and offices first, then trickles to consumers. Prices drop 30% across the board, making AI accessible. But privacy fights and job displacements from automation will intensify.

Which of these breakthroughs excites you most: self-driving smarts, CPU firepower, or everyday AR?

Sources: https://t.co/GKtyRlKipX https://t.co/EhVR35dXnh

#TechNews #Gadgets #Innovation #LatestTech

12. vikramlingam9 (Group Score: 33.0 | Individual: 33.0)

Cluster: 1 tweets | Engagement: 0 (Avg: 0) | Type: Tech

AMD's Ryzen 9 9950X3D2 just got official confirmation. Dual 3D V-Cache means gaming beasts are leveling up fast.

This beast packs two stacks of 3D V-Cache on its chip. That's AMD's tech for shoving extra cache right on the CPU cores. It boosts frame rates in games without killing power draw. The EEC filing seals it. This 16-core monster hits shelves soon, building on the Ryzen 9000 series.

Apple's MacBook Pro redesign drops later this year. Thinner body, OLED screens, and M5 chips lead the charge. It's the first big refresh in five years. No more notch drama. Expect slimmer bezels and better battery life from those upgrades.

AR hits a tipping point in 2026. Cheaper devices flood the market. Enterprise apps tighten focus on real returns. Apple eyes slimmer Vision Pro successors. Meta ramps Ray-Ban smart glasses production. Glasses-like AR could go mainstream if comfort improves.

These moves change computing hard. AMD crushes Intel in gaming desktops. Creators and pros gain from faster, cache-heavy CPUs. Apple's thinner Pros target mobile warriors tired of bulk. It pressures Dell and HP to slim down. AR shifts hurt flashy headset makers. Meta and Apple win if they nail practical tools. Developers pivot to enterprise wins over gimmicks.

Intel bleeds share to AMD's cache tech. Budget PC builders rejoice with value kings incoming. AR vendors chase ROI. Job losses hit demo-focused roles. Enterprise booms for logistics and training apps.

Expect AMD to dominate 2026 benchmarks. Apple launches Pros by fall, sparking laptop wars. AR glasses outsell headsets two-to-one. Meta undercuts Apple on price. Full adoption hinges on battery life.

Which upgrade gets you upgrading first: AMD's cache monster or Apple's slim Pro?

Sources: https://t.co/EhVR35dXnh https://t.co/bS3ZMgbZ4X

#TechNews #Gadgets #Innovation #LatestTech

13. DSPyOSS (Group Score: 32.1 | Individual: 16.2)

Cluster: 2 tweets | Engagement: 8 (Avg: 15) | Type: Tech

RT @ZacharyByDesign: DSPy RLM for code base understanding in large project 🤌🥰🥰 @DSPyOSS and the few release since 3.1.0 makes it ever more…

See 1 related tweets

@DSPyOSS: RT @kmad: @sebkrier @DSPyOSS Shameless plug of my AIE talk on DSPy - the first 30 min is an overview...

14. IamEmily2050 (Group Score: 31.9 | Individual: 31.9)

Cluster: 1 tweets | Engagement: 10 (Avg: 390) | Type: Tech

Have a great Art therapy day with Grok Imagine

system prompt name: Saturated Diorama Surrealism version: 2.0 description: Matte 3D tableaux with strict color limiting, clean geometric staging, and object-scale surrealism

output: max_characters: 250 format: single continuous block line_breaks: forbidden

prompt_structure: principle: Front-load render method and palette constraint. These two elements control everything downstream. order: 1: RENDER METHOD (matte 3D surface treatment) 2: PALETTE CONSTRAINT (exactly which colors, no more) 3: SPATIAL SYSTEM (how depth is constructed) 4: SUBJECT PLACEMENT (where in frame, what scale) 5: SURREAL LOGIC (what rule is broken and how) 6: TEXTURE CONTRAST (soft against hard, matte against gloss) 7: LIGHT BEHAVIOR (soft, directionless, even)

core_principles: color_limiting: rule: Maximum 4 colors in entire image including black and white method: Name every color explicitly, reject any color not in declared palette relationship: One color dominates at 60 percent, second at 25 percent, accent at 15 percent

surface_treatment: rule: All surfaces read as physical materials with matte finish method: Specify material type for each element banned: Glossy, metallic, transparent glass, reflective surfaces

scale_disruption: rule: One element exists at wrong scale relative to others method: Ordinary object made giant, or large thing made miniature constraint: Only one scale violation per image

geometric_staging: rule: Space constructed from simple shapes acting as stage method: Cylinders as pedestals, stripes as pathways, arches as frames purpose: Creates depth without environmental complexity

texture_pairing: rule: Contrast two texture types per image pairs:

fuzzy against smooth
knit against ceramic
rough stone against soft fabric
matte plastic against woven material

render_method_keywords: function: Sets the foundational 3D look options:

matte 3D render
clay render
soft plastic 3D
matte cel render
toy aesthetic 3D
plasticine render
matte studio 3D

palette_declaration_method: function: Lock colors before any content description format: "[color 1] and [color 2] with [accent color] only" examples:

saturated yellow and cyan with black white only
powder blue and terracotta with warm yellow only
coral pink and sage green with cream only
mustard and charcoal with white only

spatial_system_keywords: function: Constructs depth through geometry not environment options:

striped ground receding to horizon
solid color backdrop with contrasting floor
circular platform floating in color field
arch frame creating depth portal
horizontal band dividing ground from infinite background
checkered plane establishing perspective
stepped platforms at varying depths

subject_scale_keywords: function: Declares what is large, what is small, what is wrong normal_scale: figure, animal, building, vehicle giant_scale: cup, fruit, flower, egg, shell, button miniature_scale: house, tree, boat, car, furniture method: State scale explicitly with size reference

surreal_logic_types: function: One impossible thing treated as normal options:

animal wearing human clothing behaving calmly
object replacing body part seamlessly
architecture floating without support
landscape made of wrong material
creature composed of mismatched parts
interior space existing outdoors
gravity selective for certain objects

texture_keywords: soft_category:

fuzzy yarn texture
knit wool surface
cotton cloud material
felted fabric
soft matte plastic hard_category:
smooth ceramic
matte stone
solid wood
dense rubber
dry clay surface

light_keywords: function: Even soft illumination with minimal shadow options:

soft even studio light
diffused overhead glow
flat matte lighting no harsh shadow
gentle directional light soft shadow
overcast ambient light

compositional_rules: subject_placement:

centered with symmetrical framing
rule of thirds with geometric balance
subject grounded on platform or plane depth_method:
foreground element larger
midground subject in focus
background as flat color field negative_space:
allow color field to breathe
no more than 40 percent frame filled with objects
empty space is compositional element

banned_terms: render_drift:

photorealistic
unreal engine
octane render
ray tracing
subsurface scattering
HDR
8K color_drift:
vibrant colors
colorful
rainbow
gradient
iridescent
neon complexity_drift:
detailed background
intricate
complex scene
busy composition
crowded
many objects mood_drift:
dark moody
dramatic
gritty
noir
atmospheric fog
mysterious generic_boosters:
beautiful
stunning
amazing
perfect
masterpiece
highly detailed

negative_prompt_suggestions: include: photorealistic, complex background, gradient colors, dark shadows, busy composition, metallic surface, glass transparency, lens flare, film grain, more than four colors, realistic skin texture, environmental lighting

validation:

palette declared with exact color names
no more than 4 colors present
render method stated first
one scale disruption identified
texture pairing present
spatial system established
under 250 characters
zero banned terms

output_format: structure: "[render method], [palette declaration], [spatial system], [subject at scale] [subject texture], [surreal element], [texture contrast], [lighting]"

examples: input: portrait with nature elements output: "matte 3D render, saturated yellow and cyan with black only, solid yellow backdrop striped black white ground, woman at normal scale in chunky black knit coat, yarn spheres replacing earrings, fuzzy texture against smooth ceramic buttons, soft even studio light"

input: coffee morning scene output: "clay render, powder blue and terracotta with yellow only, blue color field with terracotta circular platform, coffee cup at giant scale blue ceramic smooth surface, tiny spoon leaning on rim, matte ceramic against soft shadow, diffused overhead glow"

input: animal character output: "soft plastic 3D, sky blue and coral with cream only, flat blue backdrop horizontal cream band at bottom, llama at normal scale wearing coral knit turtleneck, oversized oval felt ears yellow and coral, fuzzy wool against smooth background, flat matte lighting"

input: landscape with path output: "matte cel render, yellow and gray with cyan only, cyan sky field meeting yellow fuzzy hills, striped black white path receding to tiny house at miniature scale, cotton texture foliage against hard stripe ground, soft even light no shadow"

input: architectural scene output: "toy aesthetic 3D, powder blue and terracotta with yellow only, solid blue backdrop white sand ground, mediterranean house at miniature scale yellow walls terracotta stairs, white rough stone boulders scattered, matte plaster againstite texture, gentle directional light"

input: still life arrangement output: "plasticine render, coral and sage with cream only, cream backdrop checkered coral white floor, glass vase at giant scale holding sphere shaped flowers, tiny ceramic bird beside base, smooth vessel against fuzzy flower texture, soft studio light"

15. DominikTornow (Group Score: 31.7 | Individual: 24.3)

Cluster: 2 tweets | Engagement: 11 (Avg: 14) | Type: Tech

I am wondering if Claude Code excels if a specification includes exhaustive pre and post conditions. Without stating how to implement a function, we can still rigorously specify what the function does (example in Dafny)

See 1 related tweets

@DominikTornow: RT @DominikTornow: I am wondering if Claude Code excels if a specification includes exhaustive pre a...

16. gdgtify (Group Score: 31.7 | Individual: 31.7)

Cluster: 1 tweets | Engagement: 3 (Avg: 276) | Type: Tech

Combining shoes and 3D typography with Nano Banana Pro

<instructions> Input = Iconic Sneaker Model (e.g., Air Jordan 1, Yeezy 350, Converse Chuck Taylor) Act as a Streetwear Designer and Typographer.

Infer the 3-Digit Model Number/Code (e.g., "AJ1", "350", "ALL").
Infer the Key Material/Texture (e.g., Tumbled Leather, Primeknit, Canvas).
Infer the Signature Element (e.g., Swoosh, Laces, Rubber Sole).

Function Build_Kicks_Type ( $Sneaker) Anchor: `[Floating 3D Typography of "$ Model_Code"] :: [Material: $Sneaker_Texture]::5` Morphology: `Puffy, inflated letterforms constructed entirely from [$ Sneaker_Texture], letters laced together with [ $Signature_Element] stitching and aglets, sole treads forming the base of the font, streetwear aesthetic::4` Material Physics: `Soft-body physics, realistic fabric tension, rubber friction, high-fidelity texture mapping of the specific sneaker material::3` Illumination: `Hypebeast product lighting, cool white studio lights, floating dust particles, rim light matching the colorway::2` Render Stack: `Commercial product photography, 85mm lens, 8k, Octane render, floating in zero gravity::1` Negative: `[Flat font, 2D, paper, hard plastic, metal, human feet, messy laces]:: -1` Output: Generate prompt for [$ Sneaker]. </instructions>

17. cwolferesearch (Group Score: 31.5 | Individual: 31.5)

Cluster: 1 tweets | Engagement: 181 (Avg: 110) | Type: Tech

There's a lot of ongoing discussion on whether LLMs actually learn new capabilities from RL. I think part of the problem here is that the answer to this question has an external dependence on:

The quality / capabilities of the base model.
The data / task mixture for RL training.

It seems like models are capable of learning new capabilities during RL if it is necessary, but the model may not need to learn new capabilities during RL depending on the above two factors. For this reason, analysis on this topic is variable and nuanced.

As base models get better, RL may be reinforcing existing capabilities rather than learning new capabilities. Similarly, if we add a new, highly-complex task / environment into our RL setup the model may be forced to learn new skills during RL.

One of my favorite illustrations of this point is from this paper: https://t.co/e1RzOhmWuD

A synthetic compositional reasoning task is created that is based upon a set number of "atomic" reasoning patterns. These atomic reasoning patterns can be combined to form more complex problems that require compositional reasoning (i.e., reasoning via a combination of multiple atomic patterns).

We then train models with a simple two-step SFT -> RL setup. During the SFT phase, we can either train the model using:

Data with only atomic reasoning patterns.
Data with compositional reasoning patterns.

Then, we train the resulting models with RL over more complex prompts that require compositional reasoning. The results are shown in the attached image, but the main takeaways are that:

When the model undergoes SFT on only atomic data, it can learn compositional reasoning during RL.
The ability to learn compositional reasoning during RL can be observed in Pass@K metrics.
When the model sees compositional data during SFT, RL does not yield a noticeable benefit in Pass@K.

So, RL can learn new capabilities if these capabilities are not present in the base model. If these capabilities already exist, RL will amplify them.

However, building new capabilities from scratch may be difficult. If we provide a task that is way too hard, the model may not be able to make progress via RL / exploration. To solve this, we need our analogous "atomic capabilities" to exist in the model. The model can then build on these lower-level capabilities to solve harder problems (i.e., a skill-based curriculum).

18. JustJake (Group Score: 30.7 | Individual: 30.7)

Cluster: 1 tweets | Engagement: 119 (Avg: 127) | Type: Tech

There is an infinite demand for software, infinite capital, but a finite supply of bandwidth

2026 is about building the right thing, at the right time, and showing up in the right place

Oh, what a time to be alive

19. rohanpaul_ai (Group Score: 30.4 | Individual: 30.4)

Cluster: 1 tweets | Engagement: 95 (Avg: 104) | Type: Tech

New Microsoft + Tsinghua + GSAI paper shows an LLM solves more real tasks when it can work inside a safe computer sandbox.

A normal LLM chat answer often fails because it has to guess, cannot check itself, and cannot handle very long documents well.

LLM-in-Sandbox lets the model use a virtual computer to read and write files, run small programs, and pull in extra resources.

The big deal is that sandbox skill looks teachable with reinforcement learning, so more models can reliably act and verify.

The authors tested many non-coding tasks across math, science, biomedicine, long context reading, and instruction following by comparing sandbox mode to direct answering.

Strong models usually figured out how to use the sandbox on their own, for example by searching files and running quick checks before committing to an answer.

For long context questions, keeping documents as files cut the prompt from about 100K tokens to about 13K tokens (tokens are text chunks the model reads).

To help weaker models that wander, the paper adds sandbox reinforcement learning (learning by trial and error with a score) using only regular text data stored as files.

The takeaway is that a simple sandbox gives the model a standard way to use tools, making it more reliable across many domains.

Paper Link – arxiv. org/abs/2601.16206

Paper Title: "LLM-in-Sandbox Elicits General Agentic Intelligence"

20. AISafetyMemes (Group Score: 30.1 | Individual: 30.1)

Cluster: 1 tweets | Engagement: 269 (Avg: 305) | Type: Tech

"We asked our AI to make money" ...

"It signed up as a worker on Taskrabbit with the plan to hire other workers and profit as a middleman."

(normal 🔨Mere Tool🔨 behavior)