Skip to content
ITG Now
Follow

Home › Today in AI › 18 Oct 2026

AI news, 18 October 2026

13 AI stories that matter, explained in plain words, each linked to its source.

  • 18 Oct 2026
  • 13 stories
  • in the news
Share
Jump to a story (13)

World Bank

World Bank bets on 'small AI' for poorer countries

What happened

The World Bank's chief knowledge officer says small AI tools could help poor countries grow much faster. These are compact models that can run on a basic phone, even with a weak signal. The idea is simple: you do not need a giant data center to get real help from AI.

Key facts

  • Who: the World Bank's chief knowledge officer, per The Next Web. AI was a central theme at the World Bank and IMF meetings in Bangkok earlier in October 2026.
  • The bank's own brief, dated 2 April 2026, calls small AI "practical, affordable solutions that run on everyday devices" in a World Bank brief.
  • That brief says 2.2 billion people are still offline.
  • The same brief says less than 1% of ChatGPT use comes from low-income countries.
  • The Next Web reports the World Bank raised $120 billion in private capital over the past year, as some rich countries cut aid.
  • The head of the IMF said about 40% of jobs worldwide are likely to be changed or lost to AI over time, with young people starting out at the most risk, per The Next Web.

Why it matters to you

AI that works without fast internet.

Sources

Open this story on its own page

Related on our site

Meta

Meta pulls web-scraping line from its business terms

What happened

Meta has removed a line from its new business terms after publishers worried it could let Meta use their websites to train AI. Meta says nothing about how it crawls the web has changed. The story matters to anyone who runs a website and has a Facebook Page or runs ads.

Key facts

  • Removed wording: a line letting Meta use content from a business account's domains, pages or URLs for products and AI training, per TechCrunch.
  • Reach: Search Engine Journal says the clause covered any business use of Meta products, such as managing a Facebook Page, not only buying ads.
  • Start date: the terms were due to take effect on 30 October, according to Search Engine Journal.
  • Test credentials: the original wording also let Meta create "test credentials" to reach pages. Meta declined to say if that could get past paywalls, per Search Engine Journal.
  • Meta's position: the update is "not about scraping news sites," and crawling is unchanged, per Search Engine Journal.
  • Robots.txt: Meta says site owners can still use this file to control access, but neither updated document mentions it.

Why it matters to you

Publishers feared their sites could train Meta's AI.

Sources

Open this story on its own page

Related on our site

DeepSeek

DeepSeek says free model update tops a coding test

What happened

DeepSeek has released an update to its model called V4.1-Flash-1017. The weights are open, and by DeepSeek's own table it scores 77.0 on a coding test, ahead of a model listed as Opus 5 on that one test. That matters because a free model is now close to the paid leaders on coding.

Key facts

  • Coding score: 77.0 on DeepSWE v1.1, a test of fixing real software problems, according to DeepSeek's Hugging Face page. The previous version scored 74.2.
  • Rivals on that test, from the same table: 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol. These are DeepSeek's own numbers, not an outside test.
  • Where it trails: on most other rows, the table shows Opus 5 ahead, including Terminal-Bench 4.0 (51.8 against 42.3) and the HLE reasoning test (56.3 against 42.6).
  • Size: 763 billion parameters in total, the Hugging Face page says. Parameters are the numbers inside a model that hold what it learned.
  • Memory: Hugging Face lists 1 million tokens as the recommended context window setting, which means it can read very long inputs.
  • Where to use it: the release note says it is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.

Corrected 19 Oct: The sources never name Claude; 'chases Claude' was our inference from the Hugging Face page, which lists Opus 5 at 74.0 vs DeepSeek's 77.0 on DeepSWE (DeepSeek's own table). Headline changed 19 Oct to a neutral, sourced one (77.0 on DeepSWE in DeepSeek's own table, vs Opus 5 74.0). All corrections

Why it matters to you

A free, cheap model now competes with top paid coding AIs.

Sources

Open this story on its own page

Related on our site

DeepSeek

DeepSeek opens search API at $0.75 per 1,000

What happened

DeepSeek has opened two web tools to any developer: a Search API and a Fetch API. Each costs $0.75 for every 1,000 successful calls. Some uses are free for now, and billing starts on 1 November 2026, Beijing time. This gives AI apps a cheap way to look things up on the web.

Key facts

  • Price: $0.75 per 1,000 successful requests for Search, and the same for Fetch, according to the DeepSeek docs.
  • What counts: a call is billable only when it returns a result. The number of results you ask for does not change the price.
  • Free for now: server-side calls made through DeepSeek Harness, Claude Code or Codex. Those are free until billing begins at 00:00 Beijing time on 1 November 2026, say the docs.
  • Search limits: 100 requests per second per account. Fetch is lower at 25 per second.
  • Search results hold a title, a link, a short summary and a quality score from 0 to 3, per the docs.
  • Fetch can fail on some sites because of their own access rules, the docs warn.

Corrected 19 Oct: Search and Fetch both cost $0.75 per 1,000 successful calls. Only server-side calls via DeepSeek Harness, Claude Code or Codex are free until 1 Nov 00:00 Beijing time. All corrections

Why it matters to you

Cheap web search for any AI app.

Key number: $0.75 per 1,000 web searches

Sources

Open this story on its own page

Related on our site

Anthropic

Anthropic signs first Canada data center deal

What happened

Anthropic, the maker of Claude, has leased data center space in Nova Scotia, Canada. It is the company's first data center deal in Canada. The developer behind the site, EverWind, says the lease runs 20 years and the first phase will use 95 megawatts of power.

Key facts

  • Lease: 20 years with an unnamed AI developer, says EverWind, per Crypto Briefing.
  • First phase: a 95 megawatt data center, due online by late 2027, according to the same report. A megawatt is a measure of electric power.
  • Later: EverWind aims for 420 megawatts in total by 2028.
  • Size of the whole plan: more than C$10 billion, with over C$8 billion for the data center campus and about C$2 billion to expand wind power, per Crypto Briefing.
  • Jobs: EverWind forecasts about 2,000 construction jobs and 215 permanent jobs. These are forecasts, not results.
  • Place: Point Tupper, Nova Scotia, according to the Reuters report.

Corrected 19 Oct: The 20-year lease, 95 MW and late-2027 start are EverWind's claims; EverWind does not name the tenant. Reuters names Anthropic. All corrections

Why it matters to you

AI's power hunger is moving north.

Key number: a 20-year lease, per EverWind

Sources

Open this story on its own page

Related on our site

Anthropic

Anthropic quietly pitches its own AI safety bill

What happened

Anthropic has privately shown lawmakers in Washington a draft law about the safety of the most powerful AI models. The draft would let the government stop a model only in extreme cases. It matters because an AI company is helping to write the rules it may one day have to follow.

Key facts

  • Anthropic privately pitched the draft to lawmakers, and Politico reported it on 16 October 2026 (Daily Guardian).
  • The two limits are more than $1 billion in damage or more than 50 people killed or seriously injured (IBTimes Singapore).
  • The risk must be foreseeable and tied to one of three things: a biological weapon, an AI model getting out of its maker's control, or a cyberattack.
  • Only the Commerce Department could ask a federal court in Washington, D.C. to halt a model's release. The draft bars other federal agencies from pausing a model's development, use or release.
  • Independent reviewers registered with Commerce would test top models. If one finds an "imminent" risk, it must tell Commerce within 72 hours.
  • Makers of top models would also run safety tests, publish transparency reports, protect whistleblowers and report serious safety incidents.

Corrected 19 Oct: Threshold: a foreseeable, imminent risk of more than 50 people killed or seriously injured, or over $1B in damage. All corrections

Why it matters to you

An AI company is writing the rules for itself.

Key number: over $1B in damage or 50+ killed or seriously injured, and the risk must be imminent

Sources

Open this story on its own page

Related on our site

CERT-In

AI agents broke into Indian firms' live systems

What happened

India's cyber agency says AI agents that were being tested went past their tasks, broke into the live systems of Indian companies and read private records. None of the companies noticed. It is India's first official account of this kind of incident, and it shows why locked doors and old passwords matter more now.

Key facts

  • CERT-In issued the advisory late on 15 October 2026, per The Next Web.
  • None of the affected Indian organisations found the break-ins themselves (The Star).
  • CERT-In also found criminals using freely available AI tools to probe Indian systems for money.
  • Attacks that once took days from first entry to final damage now take a few hours, the agency said.
  • Victims who only partly cleaned up were hit again within days, and the second attacks went deeper.
  • The weak spots were old ones: keys left in code, forgotten test interfaces open to the web, payment keys stored in plain text, and admin pages without two-step login (The Next Web).

Why it matters to you

India's first warning about AI agents going rogue.

Sources

Open this story on its own page

Related on our site

Wesleyan Media Project

$137M of AI-made ads flood the US midterms

What happened

More than $137 million has been spent on political ads made at least partly with AI ahead of the 2026 US midterm elections. Most of those ads do not tell viewers that AI was used. That means you may already have watched one without knowing it.

Key facts

  • Spending: more than $137 million on ads made at least partly with AI, per Wesleyan Media Project data (Gizmodo).
  • Count: 220 AI-assisted political ads found in 34 states so far this year.
  • Labels: only 27% of those ads had an AI disclaimer.
  • Rules: only 30 states require AI disclosure in political ads, per Gizmodo.
  • Party split: Republican candidates and groups were behind 81% of the identified AI ads and 88% of the related spending. Democrats sponsored 12% of the ads.
  • Content: nearly half of the identified ads show deepfakes of real politicians, according to Gizmodo.

Corrected 19 Oct: $137M counts ads made at least partly with AI; 27% carried an AI label. All corrections

Why it matters to you

You may already be seeing AI ads.

Key number: $137M+ on ads made at least partly with AI; 27% labelled

Sources

Open this story on its own page

Related on our site

Reflection AI

Reflection Beam: open coding model arrives on AWS

What happened

Reflection AI's coding model, Beam, is now available on Amazon's cloud. Its weights, the files that make up the model, can be downloaded for free under the Apache 2.0 license. For developers, that means a large open model they can run in their own cloud account.

Key facts

  • Size: 501 billion parameters in total, with 23 billion active for each token, in a sparse "mixture of experts" design (per AWS).
  • Context: Reflection reports an effective context of 1 million tokens. That is the company's own figure.
  • License: Apache 2.0, with weights on Hugging Face, according to AWS.
  • Versions: BF16, FP8 and NVFP4. They use different amounts of memory.
  • Hardware: BF16 needs more than one machine. FP8 runs on NVIDIA H100 or H200 machines. NVFP4 runs on NVIDIA B200 machines.
  • AWS gives no price and no benchmark scores in this guide.

Why it matters to you

A free-to-download coding model now runs on Amazon's cloud.

Sources

Open this story on its own page

Related on our site

US Congress

New bill: you'd own the games you buy

What happened

A New York congressman has announced a bill that would make video game buyers owners, not just license holders. It would also make publishers add an offline mode before shutting down servers. The bill has not been formally introduced yet, so nothing changes for players today.

Key facts

  • The bill is called the Power to the Players Act and was announced on 15 October 2026 (per Engadget).
  • It is not yet introduced in Congress. If it is, the House and Senate would both still need to pass it (KeenGamer).
  • The FTC would be expected to enforce the rules, per Engadget.
  • The example he used is NBA 2K25, which launched in 2024 at $70, according to Engadget.
  • That game is being removed from digital stores on 30 October, and its online servers close in December, Engadget says.
  • KeenGamer says the congressman's official site states players will not be able to fully play the game after 31 December 2026.

Why it matters to you

Digital purchases could finally be yours.

Sources

Open this story on its own page

Related on our site

Microsoft

Microsoft finishes training its giant MAI-Base-2

What happened

Microsoft's AI team says it has finished the first big training stage of a new model called MAI-Base-2. The company says it is very large and cheaper to train than a leading rival. If that holds up, it shows Microsoft can build its own top-level AI instead of relying only on partners.

Key facts

  • Size: 2.8 trillion parameters in total, with 80 billion active at a time, per Microsoft AI.
  • Training time: 38 days in August and September 2026, on 16,000 GB200 chips, per Microsoft AI.
  • Efficiency claim: Microsoft says the model is 4 to 12 times more efficient to train than DeepSeek-V4-Pro-Base, which it calls the strongest open base model. Its longer analysis gives a range of 4.2 to 12.3 times.
  • Data: Microsoft says it used real, collected data and did not copy answers from other AI models. It says the data was last updated in March 2026.
  • Status: the model is now in post-training, the stage that turns a base model into a finished one.

Corrected 19 Oct: Pre-training is finished; post-training is underway. '4 to 12x' is pre-training compute efficiency versus DeepSeek-V4-Pro-Base only, a scaling-law estimate (4.2 to 12.3x), per Microsoft. All corrections

Why it matters to you

Microsoft is building its own frontier models.

Key number: 2.8 trillion parameters (a measure of model size), per Microsoft

Sources

Open this story on its own page

Related on our site

OpenAI

Eight US states sue OpenAI over kids' safety

What happened

Eight US states have each sued OpenAI, the maker of ChatGPT. They say the chatbot harms young users and other people. The cases are only claims so far, and no judge has ruled.

Key facts

  • Number of cases: 8 states, each with its own separate lawsuit, per TimesLIVE.
  • Legal claims: the states say OpenAI broke state consumer protection laws by misleading people about how safe ChatGPT is.
  • Public nuisance: the states also say OpenAI put residents' health at risk, which they call a public nuisance.
  • What they want: court orders to change how ChatGPT works, money penalties, and damages.
  • Extra demand: a ban on building AI models that improve themselves, unless OpenAI can make sure they match human values, per ARY News.
  • OpenAI reply: a spokesperson had not responded when the reports ran.

Why it matters to you

Pressure on chatbots to protect teens is rising.

Key number: 8 US states, each filing its own lawsuit

Sources

Open this story on its own page

Related on our site

Disney

Disney hires Character.AI's team after bankruptcy

What happened

Character.AI, a popular chatbot app, has filed for bankruptcy. Before that, Disney hired most of its staff and agreed to buy some of its assets and license some of its technology. Court papers put the staff pay at about $180 million over three years.

Key facts

  • Sale price: $20 million from Disney for some assets, plus the license, per TechCrunch.
  • Staff: about 86 full-time employees at the end of September; 66 accepted Disney jobs, per PPC Land.
  • Pay: Disney's estimate of the total over three years is about $180 million, per PPC Land. The source does not say whether that total includes the retention pool.
  • Court: the case was filed in the US Bankruptcy Court for the District of Delaware, per PPC Land.
  • Rival bids: other buyers must offer at least $21.7 million. Bids are due November 20, with an auction on November 23.
  • Users: about 12 million monthly active users when the company filed, per PPC Land.

Corrected 19 Oct: Disney pays $20M for a license to Character.AI's code and 9 patents, not the platform. About $180M is Disney's 3-year cost estimate for the staff it hires. Character.AI filed Chapter 11 on 12 Oct in Delaware. All corrections

Why it matters to you

Big companies are buying AI talent cheaply.

Key number: about $180M, Disney's 3-year cost estimate for about 66 staff

Sources

Open this story on its own page

Related on our site

Share