Omni-7 Marketing

Omni-7 Marketing Contact information, map and directions, contact form, opening hours, services, ratings, photos, videos and announcements from Omni-7 Marketing, Advertising/Marketing, Batumi.

14/05/2026

◼ Japan has the highest median age of any large country — around 49. By 2040 one in three Japanese will be over 65. The traditional model of younger family members caring for older ones is breaking down as younger people move to cities. Caregiver shortage is already hundreds of thousands of positions and growing faster than the system can produce trained workers.
◼ Three types of robots in Japanese facilities now. Physical assistants — help lift from beds, support walking. Monitoring systems — track vitals and activity. Social companions — talk, remind about medications, just stay present. Last category generates the most questions.
◼ Journalist spent several days at a facility using all three. Residents respond differently — some get attached to the companion robot, some ignore it. Staff are mostly positive because routine load shifts and there's more time for people who need a real human.
◼ The question the piece leaves open: what does it mean for a society when an elderly person talks to a robot — not because anyone decided this was ideal, but because there aren't enough people. Japan chose this answer. Not as a preference.
Source: Japan dispatch

14/05/2026

◼ Not a new attempt — UK, EU with Chat Control, now France. The argument is always the same: criminals use encryption, we need access. The counterargument is also always the same.
◼ End-to-end encryption either works or doesn't. If a service has a key to decrypt your messages — that key exists. An existing key can leak, get hacked, get obtained by people with resources. A backdoor for government is a backdoor for whoever finds it. Not an opinion — math.
◼ Signal said it'll leave if the bill passes. Not a bluff — they've exited jurisdictions before when conditions became incompatible with the architecture. WhatsApp could theoretically do the same but Meta isn't giving up the French market for a principle.
◼ Bill is still in discussion. Good time to think about which messengers you use and why.
Source: French parliament, TechCrunch

14/05/2026

◼ Noticed a pattern: AI developers and researchers started describing themselves using vocabulary from model documentation. "I've run out of context" instead of "I'm exhausted." "Need to update my weights" instead of "change a habit." "Poorly calibrated for this task" instead of "I'm not great at this." Funny at first. Then you think about it.
◼ Researchers called it LLMorphism — LLM plus metamorphosis. Not just borrowed language — it changes how you actually perceive yourself. If you habitually describe your memory as a limited buffer that can be cleared, you start treating it that way.
◼ Every dominant technology does this. 19th century through steam machinery. Then electricity. Then computers. Now language models. The difference is speed — from technology appearing to mass conceptual contamination is years now, not generations.
◼ Several hundred million people use these tools regularly. If this actually shifts self-perception at scale — that's a lot of people.
Source: academic publication arXiv: “LLMorphism: When humans come to see themselves as language models”

14/05/2026

◼ Amazon decided it wanted to measure how actively engineers use internal AI tools. Set targets, pushed them down the chain. Makes sense — they spent a lot building these things and want to see adoption. But when you tell people "we need number X," some of them think about how to get X, not how to actually use AI better.
◼ So: tokenmaxxing. Open the AI chat. Ask it to explain code you already understand. Close it. Metric moves. Ask it to write an email you won't send. Ask it anything — the point is the counter ticks. KPI is green, everyone looks fine.
◼ The funny part: every fake query costs real money. The company is literally paying for the simulation of AI adoption. So yeah.
◼ Goodhart's Law, about a hundred years old — when a measure becomes a target it stops being a good measure. The fact that this already has a stable internal name means it's not isolated incidents.
Source: internal Amazon sources

14/05/2026

◼ Billions of numbers change inside a language model while it's thinking. Concepts activating, things getting weighted, something happening. Researchers have been trying to make sense of it for years but it was essentially an unreadable mess without serious tooling.
◼ Anthropic trained a separate model — an autoencoder — to translate those internal states into short text descriptions. Kind of like subtitles for the internal monologue. Not word-for-word, but enough to see which concepts are active at a given point in the reasoning chain.
◼ Right now when a model gives you an answer you can only check the answer. Whether the reasoning that got there was actually sound — no way to tell. That's like seeing only a mathematician's final answer without the work. Readable reasoning traces change what kind of check is possible.
◼ How much of this is genuine "thought" vs artifact of the method is still open. The autoencoder is itself trained and brings its own distortions. But as a first step toward having any access to the process at all — it's a real one.
Source: Anthropic

08/05/2026

On April 29, Biohub — the nonprofit run by Mark Zuckerberg and his wife Priscilla Chan — announced a five-year Virtual Biology Initiative with a $500 million budget.

The idea: collect enough data about human cells that AI can build predictive models of how cells behave — in health and in disease. Current global datasets cover about a billion cells. Biohub wants an order of magnitude more.

$400 million goes to developing their own tools for observing and measuring biology at the cellular level. $100 million to external partners and researchers. Partners include Nvidia, Allen Institute, Human Cell Atlas, Human Protein Atlas.

Zuckerberg has talked about this before — that Biohub's long-term goal is to cure all disease through the intersection of AI and biology. Now there's money behind it.
The honest question scientists are raising: the data gap is real. But nobody knows how much more data is actually needed before the models become accurate enough to be useful in practice. Scaling worked for language and protein structure. Whether it works for cells is still an open question.

Source: Axios, TechRadar

07/05/2026

YouTube is testing a new search mode for Premium subscribers in the US. It's called Ask YouTube.

You type a question and instead of a list of videos you get a structured answer. Text plus short clips from relevant videos plus full videos alongside. You can ask follow-up questions and keep the conversation going. It's closer to ChatGPT than traditional search, except it's inside YouTube and uses real videos as sources.

The example YouTube shows: "plan a 3-day road trip from San Francisco to Santa Barbara." You get a step-by-step itinerary with embedded video clips at each point.
For viewers — easier to find the specific moment you need without watching the whole thing. For creators — more complicated. If YouTube starts surfacing clips from videos instead of full views, that hits ad revenue. There's no creator compensation policy for this format yet.

Google is clearly moving toward keeping more user questions inside its ecosystem and answering them there, without sending people to other sites. YouTube is the next step in that logic.

Source: TechCrunch, The Verge

07/05/2026

On April 28, Snapchat launched a new ad format. Brands can now place AI agents directly in Chat — the same place where you message your friends.

Not a banner. Not a video. Not a story. An agent that answers your questions, makes recommendations, and holds a conversation. Experian is the first confirmed partner in the alpha. Uber and Tinder had already tested earlier versions of the format.

The numbers Snap is using to sell this to brands: 950 billion chats in Q1 2026. 85% of users open Chat regularly. Over half a billion people have messaged My AI since its 2023 launch. Existing Sponsored Snaps already drive 22% more conversions at nearly 20% lower cost per action.

The logic is clear — people in chat are already in conversation mode, already open to interaction. The question is how ready they are for the other side to be a brand agent rather than a friend.

For marketers this is an interesting format. Instead of guessing your audience in advance, personalization happens live inside the conversation based on what the person actually says.

Source: TechCrunch, Snap Newsroom

06/05/2026

If you use Chrome and your disk has been mysteriously shrinking lately — go into your Chrome profile folder and look for a directory called OptGuideOnDeviceModel. There's likely a file called weights.bin inside. Around 4GB.

That's Gemini Nano — Google's on-device AI model. Chrome installed it between April 20 and 29, on hundreds of millions of devices worldwide. No dialog, no checkbox, no notification. It just decided your disk was a reasonable place to store it.

If you delete it, Chrome treats that as a system error and re-downloads it the next time its configuration server tells it to. The architecture doesn't register "user deleted this" as an intentional choice.

There's also this: Chrome 147 has an "AI Mode" button right in the address bar. Looks local — you've got the model installed after all. But every query you type into it goes to Google's servers. The on-device model powers other features, not that button.

Privacy researcher Alexander Hanff published forensic documentation of the whole thing. Under Europe's ePrivacy Directive, storing anything on a user's device without explicit consent is illegal. Chrome is doing this across the entire EU.

Source: That Privacy Guy, CyberNews, Tom's Guide

06/05/2026

On April 24, DeepSeek released V4 — and again made everyone reassess what's actually happening in the AI model market.

V4 Pro on coding benchmarks sits alongside GPT-5.4 and Claude Opus 4.7. Highest Codeforces rating of any model at release. On SWE-bench Verified — within 0.2 points of Claude Opus 4.6. API price: $3.48 per million tokens versus $25 for Claude. Open weights, MIT license, self-hostable without contacting DeepSeek.

One separate detail: V4 was trained on Chinese hardware — Huawei Ascend 950 chips, not Nvidia. First time a model at this level has been built entirely outside the Nvidia ecosystem. Jensen Huang called this outcome "horrible for the United States." Pretty obvious why.

NIST evaluated that V4 trails leading closed models by about 8 months — but the price gap is in the tens of times. For most real-world tasks that tradeoff is starting to look reasonable.

Source: TechCrunch, DataCamp, NIST, Codersera

Address

Batumi

Telephone

+995595012669

Website

Alerts

Be the first to know and let us send you an email when Omni-7 Marketing posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Contact The Business

Send a message to Omni-7 Marketing:

Shortcuts

Share

Category