How AI Helps
Open in Telegram
How artificial intelligence helps people and teams at work and at home. Short, sourced briefs on AI agents, automation, tools, workflows, and business use cases: what happened, why it matters, and how to apply it. https://t.me/howaihelps?direct
Show moreThe country is not specifiedThe category is not specified
599
Subscribers
No data24 hours
-27 days
+28430 days
Data loading in progress...
Similar Channels
No data
Any problems? Please refresh the page or contact our support manager.
Tags Cloud
No data
Any problems? Please refresh the page or contact our support manager.
Incoming and Outgoing Mentions
---
---
---
---
---
---
Attracting Subscribers
July '26
July '26
+25
in 1 channels
June '26
+598
in 21 channels
Get PRO
May '26
+17
in 0 channels
| Date | Subscriber Growth | Mentions | Channels | |
| 14 July | +1 | |||
| 13 July | +3 | |||
| 12 July | +3 | |||
| 11 July | +1 | |||
| 10 July | +3 | |||
| 09 July | 0 | |||
| 08 July | +1 | |||
| 07 July | +1 | |||
| 06 July | +2 | |||
| 05 July | +5 | |||
| 04 July | +1 | |||
| 03 July | +2 | |||
| 02 July | +1 | |||
| 01 July | +1 |
Channel Posts
South Korea makes AI access eligible for public support
South Korea's Cabinet approved rules for national and local governments to help defined AI-vulnerable groups pay for AI products or services. The decree takes effect on 21 July.
The official Cabinet briefing confirms the amendment. Reported groups include disabled people, those aged 65 or older, basic-livelihood and near-poor recipients, and job seekers.
This is not automatic free AI access. Support depends on budgets, and each authority must publish its criteria and application process. No subsidy amount is stated.
Vendors get an AI-product confirmation route through KOSA and TTA review. Confirmed products are due to receive public procurement advantages from next month.
The policy acts on both sides: people may get help with AI costs, while verified vendors gain a government purchasing route.
| 2 | The AI asset worth owning is not your promptāit is the test set that stops your team from learning the same failure twice
The new model is faster and cheaper. It also scores higher on public benchmarks. In the demo, it writes cleaner support replies. The team switches.
Two weeks later, a customer asks for a refund after receiving a replacement shipment. The model sees the general 30-day policy and confidently approves it. The real rule is different: a person must first reconcile the replacement status. The old system made this exact mistake once. Someone corrected it in a chat, then the lesson vanished with the session.
This is not just a model regression. It is a memory failure.
Teams often protect a clever prompt while treating corrections in tickets and reviewer comments as disposable QA. Yet prompts change with models and interfaces. A reviewed set of real failure cases can test the next systemāfrom its prompt and model to its retrieval setup or provider.
Imagine two teams evaluating an upgrade. One compares three polished demos. The other reruns 300 moments when its AI misunderstood a policy; cited a report that did not support the claim; turned missing data into zero; or completed a CRM update while creating a duplicate. The second team can still choose the new model. It also knows what that choice costs.
An eval suite is executable institutional memory.
Build it as a cycle:
real failure ā human correction ā reusable case ā model change ā regression signal ā policy review
For every meaningful correction, create one acceptance case. Record the situation; the behavior that must hold; the evidence of a pass; and the source and review date of that expectation. Keep hard checks separate from judgment calls. A citation can be verified mechanically; fairness or a justified exception may need human review.
Store the smallest authorized version of the incident that still preserves the tested boundary. A rare or sensitive case may need to stay restricted or be recreated as a synthetic scenario.
The return arrow matters too. A stale case can preserve an obsolete policy; a narrow customer population; or old bias. The goal is governable memory, not sacred history.
Your model name and pricing may change. So may the context window and vendor dashboard. The acceptance cases should survive them.
The best AI teams may not have the cleverest prompt. They may be the teams that never have to learn the same failure twice. | 34 |
| 3 | The battery bottleneck in an AI gadget may be its answer: each extra word can cost more than another input token
You point repair glasses at a bicycle wheel. One prompt asks, āDescribe everything you see.ā Another asks, āMain part? Three words.ā The image and model stay the same, yet the battery bill may not.
A July 2026 study tested five quantized vision-language models from 1B to 4B parameters on an RTX 3070 laptop GPU and a Jetson Orin NX. In those specific setups, each generated output token added 11ā39 times as much wall-clock time as an input token. In one 448Ć448 RTX 3070 comparison, sequential decoding accounted for 86ā97% of total energy.
The reason is a narrow doorway. The model can process the image and prompt together during prefill. It then writes the answer one token at a time during decoding. In the study, more complex images produced up to 4.1Ć energy differences at the same resolution because they triggered longer descriptions.
The answer box is part of the battery architecture.
This does not mean every answer should be tiny. āRear axle bolt | confidence 0.82 | verifyā may be enough for a low-risk identification task. The same compression can hide an essential warning when the view is unclear.
A better design is progressive disclosure with three modes: FAST gives the smallest useful fact, EXPLAIN adds detail on demand, and STOP appears when uncertainty or risk cannot fit safely into a short reply.
Here is a compact response contract for a prototype:
You are the perception layer of a low-risk prototype.
Do not issue physical commands.
FAST mode:
label: [maximum 3 words]
confidence: [0.00-1.00]
review_needed: [true/false]
If the fact is not visible, the image is ambiguous,
confidence is below [threshold], or a safety-relevant
detail may be present, return:
label: cannot determine
confidence: [value]
review_needed: true
Use EXPLAIN only when requested. Give visible evidence,
alternatives, and limits in at most [N] tokens.
To test it, run the same image and task with 8, 32, and 128 output-token caps. Record the actual token count, total latency, correctness, preserved uncertainty, and preserved stop condition. Repeat each run five times after warm-up. If you lack a calibrated energy sensor, report tokens and latency rather than turning a battery percentage into an energy claim.
The efficient visual assistant may not be the one that sees less. It may be the one that knows when three words are enoughāand when they are not. | 42 |
| 4 | Chat was built for answers, but agents need semantic diffs when they change a forecast, calendar, presentation, or customer database
A finance lead asks an agent to refresh the quarterly forecast from new sales data. Forty minutes later, it returns a polished workbook: āUpdated assumptions, reconciled totals, and refreshed charts.ā
Did revenue improve, or did the agent quietly change what counts as likely revenue?
That cheerful summary hides every question that matters. Were formulas changed, or only inputs? Did a missing region become zero, stay blank, or get inferred from last quarter? Can the old forecast be restored without discarding later human edits?
More prose will not solve this. A story about the work forces the reviewer to reconstruct state from the agentās description of state.
Software already has a stronger object: the pull request. Work happens in an isolated proposal, changes are inspectable, checks run before acceptance, comments stay anchored, approval is explicit, and history survives the final summary.
The important invention is not Git. It is the separation between generating a change and making that change real.
Office agents need the same contract, expressed in the language of the work. A raw comparison might say:
F27 changed from 0.18 to 0.21.
A semantic diff would say:
Europeās conversion assumption rose from 18% to 21%, based on the CRM snapshot from July 12. Projected revenue increased by +ā¬420k. Two charts and one hiring threshold are affected.
The second version shows changed meaning, supporting evidence, and downstream consequence. It compresses hundreds of recalculated cells into the few decisions a responsible person must inspect.
The same pattern changes with the domain. A calendar diff should show whose time moved and which constraints broke. A presentation diff should surface a changed claim, source, chart scale, or hidden caveat. A CRM diff should separate added facts from overwritten human judgment and show which automation will fire.
AI can make a thousand coherent edits in the time a person needs to understand ten. The interface must compress the review burden without compressing away the risky consequence.
A useful test for any agent product:
ā¢ Can I inspect the proposal without accepting it?
ā¢ Does the comparison show changed meaning and downstream effects, not only changed cells?
ā¢ Could a second person later reconstruct who accepted which evidence?
A semantic diff is a decision surface, not proof. Sending an email to 20,000 customers is not reversible. The diff can expose recipients, claims, and segmentation before sending, but it cannot manufacture an undo button.
The trustworthy handoff is no longer āI finished.ā It is: āHere is the smallest accurate picture of the world I propose changing.ā
The future of agent interfaces may look less like better conversation and more like a well-designed difference. | 45 |
| 5 |  Thomson Reuters cuts engineers as it hires for AI
Thomson Reuters confirmed cuts as it deploys AI in legal, tax, and regulatory products. An employee said up to 500 jobs could go, but the company called it a āsmall numberā and did not confirm the figure.
The Reuters report says it plans more than 250 net-new engineering roles worldwide over two years, mostly senior and āAI-nativeā. These are planned hires; locations and timing of cuts are undisclosed.
This is more precise than āAI replaces engineersā. The company is removing roles while seeking people who can combine models, proprietary data, and judgment in regulated workflows. At 500 jobs, the cuts would equal 5.2% of its 9,400-person operations and technology unit.
For engineers, coding skill alone may not be enough. Production judgment in high-stakes AI products matters more in this hiring mix. | 46 |
| 6 |  Canadaās bank supervisor flags vendor AI risk
Canadaās bank regulator wants banks and insurers to map AI inside outsourced services. Its playbook says to connect those dependencies to critical operations, test failures, and keep manual fallbacks.
In its technology-risk bulletin, OSFI says firms should require vendors to disclose whether and how they use AI to deliver services, and assess replacement options.
This widens the AI inventory beyond tools a company buys directly. Procurement, security, risk, and operations teams need to find AI embedded by suppliers before an outage or model withdrawal disrupts critical work.
The bulletin sets non-binding sound practices, not regulatory expectations. The useful test is simple: if a supplierās AI disappears, can the business still run? | 45 |
| 7 | The most convincing synthetic family history may not look like a fake ā it may look like the bedtime story everyone loves
A child tells the family how a great-grandparent crossed town with a red umbrella to save the grocery shop.
The adults look at one another. Everyone knows the scene. Nobody remembers it happening.
Weeks earlier, AI had turned a few real fragments into a bedtime adventure: an old shop photo, a recipe card, one memory of hiding under the counter, and a disagreement about the red awning. The model supplied a rainy Tuesday. Then came the umbrella. A missing delivery created the problem. The perfect joke became the ending.
The story was not created to deceive anyone. The material began with love.
That is exactly why the detail became persuasive. Everything around the umbrella was real, so the invention borrowed credibility from its neighbors. Each retelling made the scene easier to remember. An illustration gave it a fixed shape. Familiarity slowly started to feel like proof.
Families have always embellished stories and disagreed about the past. AI gives the process new speed and polish. It also makes the result unusually stable. A persistent family storyteller can repeat the same invented detail in later episodes and images. It may even place it in a printed family book. The version with the fewest gaps may become the version a child inherits.
The useful boundary is simple: leave the seam visible.
Keep four kinds of sentence separate:
Grandma remembers this.
The photo shows this.
The family disagrees about this.
We invented this part.
A child can enjoy all four. The risk begins when generated coherence is allowed to speak with the authority of a witness.
This also matters when family history is painful or disputed. A smooth childrenās version may quietly favor the relative with the largest digital archive. It may replace conflict with a cleaner motive that suits the story.
Let the red umbrella stay. Let it become a beloved symbol. Just keep the sentence that says where it came from.
A family story may deserve to survive without pretending every part of it happened. | 50 |
| 8 |  Intel puts ā¬5 billion behind Europeās AI chip supply
Intel will spend ā¬5 billion at its Leixlip campus in Ireland to make more Intel 3 wafers for Xeon 6 and next-generation Xeon processors used in AI. Work began in 2026, with most spending expected by the end of 2027.
Intelās official announcement says the project will upgrade existing fabs, add equipment and link campus modules with a larger automated track system. It will use current cleanroom space, not build a new fab.
Leixlip is Intelās only site running Intel 3. The expansion is set to add several hundred permanent high-tech jobs and retrain staff.
For AI buyers, this could mean more advanced server chips made in Europe. Intel gave no figure for added capacity or customer allocation, and the ramp depends on demand, labour, equipment and execution.
Scaling AI also needs factories and skilled people. | 58 |
| 9 | AI-generated screens can remove menus and friction, yet they also erase the shared evidence people need to support, audit, or challenge software
A customer calls support: āI clicked the blue review button.ā The support agent replies: āThere is no blue review button.ā
Both are right.
The app assembled that screen for one person using the current data and permissions, then shaped it around inferred intent. After the action, the screen vanished. Another user might receive a summary card or no review option at all.
This is the hidden shift behind generative interfaces. A static screen was not only navigation. It was social infrastructure. A colleague could point to the same control. Documentation could name a path. Support could reproduce a problem. A reviewer could compare what two people were offered.
AI can now stream validated components and data into a native interface built around the task. That can be much better than forcing everyone through the same menus. The screen may feel unusually clear to its user while becoming invisible to everyone who must understand it later.
For a consequential moment ā a financial decision, a health action, or a change in access ā the product needs a Screen Receipt. This is a compact record of the decision surface, not a recording of every click.
It should preserve four things:
ā¢ State: what was represented, when, and with what uncertainty.
ā¢ Rights: which actions were allowed or omitted, plus the permission basis.
ā¢ Claims: which warning or recommendation shaped the choice.
ā¢ Recovery: a canonical view another person can inspect without replaying the original model run.
This suggests a useful design rule: keep a stable public grammar and let AI generate the private composition. The team fixes component meanings and permission rules. AI can adapt the arrangement and emphasis to the task.
The receipt should minimize personal data. Preserve what is necessary to reconstruct the choice, not the full conversation. Inferred personality should never alter a consequential price or warning. The same boundary applies to consent language and appeal rights.
When evaluating a generative interface, ask two questions. Can a second person reconstruct what the user saw? Can they explain why it appeared and what action the user believed they were taking?
Trustworthy variation is more valuable than infinite variation.
The screen can be temporary. The meaning of the screen cannot be. | 54 |
| 10 | A beautiful AI image can start with one prompt, but a meaningful visual story starts with twenty real objects and a question
You are making a 30-second projection for a school show. Your first prompt is āfuturistic and dreamlikeā. The result looks polished and could belong to anyone.
The Smithsonian Dreams project uses a custom AI system to draw connections across millions of digitized Smithsonian objects and records. Those connections become an evolving projection across the Castle. The useful lesson is not the scale or the visual style.
The collection is the modelās world. Curation decides what that world is allowed to say.
Try a Twenty-Object Dream. Choose one question: āHow did humans carry a message before a smartphone?ā Then gather twenty items from your own material or from sources with permission and a clear reuse license.
Before generating anything, give each item a provenance line: title, date, exact creator or culture wording, source URL, object ID, rights marker, and one verified fact. That line is your Collection Provenance Sheet.
A letter can lead to a signal flag, then a telegraph key. From there, a pager can meet a satellite component. The sequence might become physical transport ā encoded signal ā invisible network.
AI can propose bridges across supplied dates and functions. It must keep resemblance separate from documented influence. A smooth invented story is still an invented story.
Paste your twenty records into this prompt:
I am building a visual story from 20 items.
Question: [insert one question]
Format: [choose one final medium]
Records: [paste these fields for every item: title | date | creator or culture | source URL | object ID | rights marker | one verified fact]
Act as a source-bound collection assistant, not an art generator yet.
Use only the supplied records.
Do not invent any unsupported relationship or meaning.
Label each relationship as documented / shared metadata / visual resemblance / speculation.
Keep the source URL, object ID, and rights marker attached to every item.
Flag gaps instead of filling them.
Return:
1. Five possible theses with exact supporting facts.
2. Three pairings for me to verify.
3. Items that do not fit.
4. A three-act storyboard for the strongest thesis.
Do not generate images yet.
Choose one thesis and arrange it as setup ā collision ā changed view. Keep object IDs and source captions visible in the rough storyboard.
Then run the remove the AI gloss test. If the thumbnail sequence remains interesting without effects, the collection has an idea. If it collapses, you have a mood board.
Only then let AI build transitions and help shape the final medium. When using museum records, read the rights marker on each object. CC0 can clear copyright held by the institution without clearing privacy or cultural restrictions.
Twenty objects will not match Smithsonian scale. They do something more useful for a small creator: they give the model enough reality that it cannot replace your point of view with a mood.
AI can animate a collection, but only a curator can decide what the collection is allowed to say. | 60 |
| 11 |  AI disinformation is now an assembly line
CNA examined 30 TikTok accounts and more than 550 videos targeting Singapore and Malaysia. Female presenters made repeated claims look independent; the clips drew over 3 million views.
The investigation found 98% used AI-generated, manipulated, or copied female personas, often reusing voices or scripts. Nearly nine in 10 pushed false or misleading claims. TikTok removed two flagged accounts for deceptive behaviour.
The upgrade is a workflow, not one flawless deepfake. Many faces, scripts, and shared audio can make one narrative feel like consensus.
Checking one face for AI is no longer enough. Viewers and newsrooms must compare accounts for repeated wording, reused audio, and unsupported claims. CNA could not identify the operator or tell if the effort was commercial or state-linked; views show reach, not persuasion. | 51 |
| 12 |  AI demand is now visible in TSMC's sales ledger
TSMC reported June revenue of NT$442.68 billion, up 6.2% from May and 67.9% from a year earlier. This lifted AprilāJune sales to a record NT$1.27 trillion, or $39.62 billion, 36% above last year and slightly above analysts' estimate.
The official June sales release gives a hard measure behind the AI infrastructure boom: the world's largest chip foundry is turning demand into revenue, not forecasts. As a supplier to Nvidia and Apple, TSMC's numbers matter to chip designers, cloud operators, and equipment makers.
For accelerator supply, data center capacity, and semiconductor investment, planners can see leading-edge demand kept expanding through June.
One limit matters: the monthly data is unaudited and does not split AI from non-AI sales. It reports revenue, not profit. Product mix, margins, and management's demand view are due with Q2 earnings on 16 July. | 62 |
| 13 | Your next AI gadget may be only a link: the web page is its casing, but the visitorās weakest device defines the experience
You open a webcam demo on your laptop. Your hand moves toward the camera, and the flat video becomes a responsive 3D point cloud. There is no inference API waiting on a remote server. The model is running inside the browser tab.
Googleās LiteRT.js release makes this shift tangible. A small .tflite model can download once, prepare tensors in JavaScript, then run through CPU/Wasm or WebGPU when available. The same link can power depth, object detection, vector search, or image upscaling.
The web page is the casing. The visitorās chip is the motor.
This changes the economics and the feel of a small AI project. Live input can stay close to the user, and there is no per-inference server bill. A browser tab starts to feel like a gadget you can hand to someone as a URL.
Then comes the Chromebook surprise.
On a recent laptop, the point cloud feels instant. On an older machine, the first load takes too long and the frame rate collapses. On a browser without WebGPU, the page hangs because nobody designed a CPU fallback.
Browser AI moves the server problem into product design. You no longer control the inference machine; every visitor brings one. A 150 MB model is not instant because inference becomes fast after loading. āNo cloud inferenceā does not mean āworks offlineā. Local execution is a claim to test, not a privacy label.
Before publishing, fill in this Browser AI Reality Card on two devices:
BROWSER AI REALITY CARD
Project claim:
Model / version / source:
License checked:
Model download size:
Device / browser / backend:
First-load time / median speed:
No WebGPU fallback:
Low-memory or denied-permission path:
What input enters the model:
What leaves the device:
Resources still fetched remotely:
Works after load with network off:
Start with fake input and prove correctness on CPU/Wasm. Connect the camera or files only after the output makes sense. Add WebGPU as acceleration, not as the only path. Inspect network requests after loading and during inference. A page may infer locally while still calling analytics or fetching third-party resources.
The page is now the casing. The visitorās chip is the motor. Your real job is to say exactly what the device does on a machine that is not yours. | 57 |
| 14 | Two robots can both claim 90% autonomy, while one needs ten seconds of help and the other needs a technician for twenty minutes
Nine objects land in the tray. On the tenth, a flexible bag collapses under the gripper and blocks the work area.
A worker leaves another station to clear the bag and rebuild the scene before restarting the cell. Is this still 90% autonomy?
That number hides the robotās real job: recovering the room after something goes wrong.
Recent robotics research is starting to separate three outcomes:
ā¢ Retry: The goal and scene are still valid. The robot can attempt another path.
ā¢ Reset: The failure changed the scene. The robot must restore a valid starting state before work continues.
ā¢ Handoff: Recovery exceeds the robotās skill, permission, evidence, or risk boundary. A named human role takes over.
FLARE, a 2026 failure-aware framework, makes this distinction explicit. It uses failure videos and a library of object-centered reset skills to return a scene to a usable state. Other recent work studies safe resets without human intervention and the cost of querying a person during robot recovery.
These systems do not mean general recovery is solved. They show a more important shift: reset is becoming a capability, not an awkward scene edited out of the demo.
This gives buyers, workers, and anyone watching a robot demo a better measure: recovery-adjusted autonomy.
Two robots may complete nine out of ten cycles. Robot A pauses once an hour and asks a nearby worker to rotate a box. The interruption takes ten seconds. Robot B jams a soft package beneath a conveyor guard and needs a technician for twenty minutes.
The benchmark calls them equal. The workplace does not.
Use this quick audit when someone presents an autonomy percentage:
ā¢ How often does help become necessary?
ā¢ How long until useful work resumes?
ā¢ What expertise and physical access are required?
ā¢ What damage or risk exists before the stop?
ā¢ Whose planned work gets interrupted?
Human help is not automatically a flaw. A clean handoff can make a robot safer and useful sooner. Hidden handoffs create intervention debt: the machine appears autonomous because someone else quietly restores its world.
A useful robot does not need to be invincible. It needs to know whether to try again, repair the scene, or call the right person without making their job harder.
The demo shows the task. The deployment begins with what happens next. | 72 |
| 15 |  Hyundaiās robot plans enter the labor contract
Hyundai workers in South Korea began two-hour partial strikes from July 13 to 15 after wage talks failed. Pay and bonuses are the main dispute, but AI-related job and working-condition guarantees are also on the table.
A Bloomberg report says the walkouts involve a union of roughly 40,000 workers. Hyundai plans to deploy Atlas humanoids from Boston Dynamics at its Georgia factory in 2028, though Atlas is not yet working on Korean lines.
Workers are bargaining over AI before the machines reach their jobs, not after replacement begins. The question is who gets job security and under what conditions automation proceeds.
For manufacturers, physical AI is now a labor issue as well as an engineering plan. Consultation, retraining, pay, and deployment rules may need to be settled before robots can scale. | 71 |
| 16 | An AI agent can now write its own temporary loop, so the real builder skill is designing the fence around what that loop may do
A teen gives an asset scout twenty candidates for a game jam. Instead of discussing every file one by one, the agent writes a tiny JavaScript program. It checks metadata and licenses in parallel, rejects incomplete records, compares formats with project requirements, and returns six candidates with the evidence behind each result.
Those twelve lines are not the product. They are ephemeral workflow code: a small program written for one run to coordinate allowed tools.
GPT-5.6 made this pattern explicit with programmatic tool calling. Repetition, joins, filtering, arithmetic, retries, and parallel lookups can happen inside code instead of filling the model's context with twenty bulky tool responses.
The useful frame is:
Code handles the repeatable middle.
The model handles the ambiguous edge.
The human owns the consequential action.
The tension is easy to miss. A compact six-row result can be faster and cheaper to judge. It can also hide the missing license, repeat a failed call twenty times, or compress away the source needed for verification.
Boundedness creates the gain. Before an agent writes its loop, give it a Tool Choreography Card:
Design one bounded programmatic-tool stage for this task.
Goal: [one mechanical result]
Allowed read-only tools: [names and exact input/output schemas]
Maximum records: [N]
Maximum calls per tool: [N]
Calls safe to run in parallel: [list]
Required result: [exact JSON shape]
Evidence to preserve: [sources, licenses, rejection reasons, errors]
Retry rule: [zero or one retry]
Stop condition: [exact point]
Never guess missing values or repeat a completed call. Return missing fields and failures as structured unresolved items. Do not include writes, purchases, posts, messages, uploads, downloads, deletions, or account actions.
After the program returns, use model judgment only for: [ambiguous comparison]
Require human approval before: [every side effect]
Return the card, bounded pseudocode, three failure cases, and a fake-data test.
For the asset scout, code can prove format compatibility and preserve the creator, source, license, and exact rejection reason. The model can compare the remaining art styles. Downloading or publishing the winner stays beyond the loop.
The future agent may write its own loop. You still write the fence around the loop. | 78 |
| 17 | AI video can make every shot look beautiful while forgetting the one fact your story needs when it returns from off-screen
A paper robot enters a blue bedroom with a red key in its left hand. It puts the key inside a clear jar and closes the lid.
The camera follows the robot into the hallway. When it returns, the light has changed from warm to cool. The robot opens the jar ā and finds two keys. Or the jar has moved. Or the key is suddenly in its right hand.
Every frame may look convincing on its own. The sequence is still wrong.
This is the hard part of longer AI video: continuity becomes visible when the story asks the model to remember something it cannot currently see.
At the ICML 2026 From Frames to Stories workshop on July 10, researchers organized this challenge around persistent state, creator control, and evaluation. The direction matters: making a longer video is no longer just a matter of writing a longer prompt. Creators need to manage what remains true between shots.
A story is a memory test disguised as cinema.
A reference sheet says what the robot and key look like. A beat sheet says what happens next. A Story State Packet records what is true now: the key is inside the jar; the lid is closed; the jar is on the desk.
Use this small version before generating a sequence:
STORY STATE PACKET
Three facts that must persist:
1.
2.
3.
Two changes that are allowed:
1.
2.
One event that changes the state:
One object that goes off-screen:
Return shot that tests its state:
Forbidden drift:
Last shot verified by a human:
Then run a Five-Shot Drift Test:
ā¢ Shot 1 establishes three persistent facts.
ā¢ Shot 2 visibly changes one object or relation.
ā¢ Shot 3 moves that object off-screen.
ā¢ Shot 4 changes lighting or camera angle without changing story state.
ā¢ Shot 5 returns to the object and requires its new state to persist.
Generate one shot at a time. Update the packet only with facts visible in the accepted output, not facts that existed only in your prompt. Score continuity separately from visual quality. If the sleeve became wet in shot 1 and is dry in shot 3, prettier lighting does not repair the missing event consequence.
This turns drift into evidence. You can tell whether the workflow lost appearance or state, and whether the break is spatial or causal. Then restate a constraint or regenerate one shot. You may also edit around the failure or deliberately allow the change as part of the style.
A longer video is not more frames. It is more promises to remember. | 73 |
| 18 | An AI agent turned Telegramās open source into a personal app in one day ā the scarce part was no longer code
I opened Telegram and saw the same old choice: hundreds of unread channel posts, or one āmark all as readā action that hides the useful ones.
I wanted a third option: choose a few channels and read all their unread posts in one chronological feed.
That feature did not exist, so I described the behavior to Claude, connected a Redmi Note 8 Pro over USB, and used my own Telegram build that evening.
The agent moved the project to the latest published Telegram source, understood its local database and navigation, built the APK, installed it, opened it on the phone, inspected the screen, and iterated.
The result was a new Digest tab with channel selection, a separate reading boundary for each account, infinite scrolling, albums, reactions, and Telegramās native interface.
The first version still had scroll jitter.
I did not tell the agent which class to edit. I described the visible defect and asked it to reproduce, measure, fix, reinstall, and measure again. Janky frames fell from 11.9% to 2.1%.
The new frame: source code is no longer a finished product. It is raw material for software shaped around one person.
This works best when three things are true: the source is open, the desired behavior lives on the device, and the result can be checked directly.
When requesting a personal feature, describe outcomes instead of code:
Problem: [what I do manually or what is missing]
Desired behavior: [what I should see after each action]
Data and state: [what must be stored and survive a restart]
Boundaries: [what must not change]
Done when:
- [verifiable criterion 1]
- [verifiable criterion 2]
- [verifiable criterion 3]
Reuse the appās existing UI and architecture. Implement the smallest complete version, build it, install it on the connected phone, and verify every criterion on the device.
The agent can handle the repository, toolchain, build, installation, logs, and repeated fixes. Your highest-value work moves upstream: define the experience precisely, protect local credentials, and decide what ādoneā looks like.
Read the full walkthrough and use the ready prompts. | 77 |
| 19 | We tested two new agents on the same visual IQ test ā GPT-5.6 SOL was the fastest, while Claude Fable 5 nearly closed Claudeās old gap
Two new agents have now clicked through the same 25 tiny visual puzzles from post 303.
Claude Fable 5 scored IQ 121 in 57 minutes 14 seconds. That is the strongest Claude result in the experiment, 31 points above the earlier Claude runs.
Codex Ā· GPT-5.6 SOL scored IQ 126 in 8 minutes 45 seconds. It finished faster than every previous run and beat the comparable Codex 5.5 result by two points.
The numbers are interesting, but the methods explain more.
Fable followed the direct visual route. It worked through the questions in order, downloaded unclear images at full resolution, and even measured pixel brightness and dot positions when normal zoom was not enough.
SOL used a hybrid agent workflow. It solved the opening questions in the live browser, then combined visual checks, parallel agents, earlier test materials, answer-key research, and browser automation. That made it remarkably fast, but it also means IQ 126 is an agent-orchestration result, not a clean blind model score.
Codex 5.5 still holds the absolute record at IQ 131, though that older run used a shorter prompt and the siteās default age. The practical lesson has not changed: on visual web tasks, the workflow can matter as much as the model.
See the updated table, screenshots, exact methods, and result links | 77 |
| 20 |  Apple reportedly plans 1.5 TB for local AI
Apple is reportedly designing its M7 Ultra chip with up to 1.5 TB of unified memory. That is twice the maximum being considered for the M5 Ultra and points to a new class of local AI workstation.
In its report, Bloomberg says the top configuration will depend on industry conditions. Apple has not announced the chip, and the roadmap can still change.
Unified memory is shared by the CPU and GPU, so its capacity affects how much model data the system can hold. If shipped, this Mac could let developers and privacy-sensitive teams keep larger model weights and datasets on one machine, reducing cloud use for workloads that fit. Memory alone does not establish inference speed.
Apple may be targeting the next bottleneck: enough shared memory to run serious AI workloads privately on one workstation. | 72 |
