Using Local Coding Agents
Short note linking a new article on setting up local coding agents with open-weight models.
Latest AI/ML News
35407 matching items
A self-sustaining private island in Canada sold for $6 million — it can only be reached by boat or helicopter
A luxurious lodge on a private island in Nova Scotia with enough infrastructure to maintain life off the mainland was sold for $6 million.
How to make Linux look like MacOS for free - with a few simple ZorinOS tweaks
Looking to migrate to Linux but fond of the Mac UI? Zorin OS can help make your new distro look very much like the one you just left. Here's how to achieve that for free.
When software developers and AI agents share the learning
Before Tobi Lütke ran Shopify, he learned programming through Germany’s apprenticeship system, the way people have learned trades forever: in a shared workshop, watching people who already knew what they were doing. More recently, describing Shopify’s River , he reached for a related word: Lehrwerkstatt , a teaching workshop where “the whole shop floor is the classroom.” X has been agog by the numbers around River , Shopify’s Slack-native AI agent . In total, 5,938 Shopify employees worked with River across 4,450 different Slack channels, and River now coauthors roughly one in eight merged pull requests across the company. It’s a big deal, but understanding why it works that way is the most important part. River can read code, run tests, open pull requests, query the data warehouse, inspect production traces, and sometimes push back on a plan it thinks is bad. Great. Lots of companies will have clever coding agents someday soon. Some already do. The interesting part is that River doesn’t work alone; it works where everyone can see it. Betting on the workshop I’ve already argued that agents reward explicit, consistent, well-documented software . They like the “boring” stuff, such as schemas, tests, conventions, clean setup instructions, and codebases that don’t require a deep retrospective with the one engineer who remembers why the build script has to run twice. Dropping an agent into a messy repo is mostly an efficient audit of your engineering discipline. Agents hold up…
AI needs a flight school
In the late 1960s, elite Navy pilots began losing dogfights. The deep, instrument-level understanding of exactly where they were, what their aircraft was doing, and what was coming next had been automated. And when moments of crisis arrived, they didn’t have the situational awareness to respond. Put a plane on autopilot long enough, and the pilot stops actually flying. The same dynamic is playing out across enterprise software. AI is generating code faster than developers can understand it , and leaders are celebrating the velocity without asking who’s actually flying the plane. A developer who has only ever “vibe coded” has perception at best. They can “see” the outputs but can’t fix any internal failures caused by the very AI systems they’re relying on. The easiest thing to do is to say the answer looks good enough. Cut and paste it in and hope it works out. According to Model Evaluation & Threat Research’s randomized control trials , experienced developers working with AI tools actually took 19% longer to complete tasks than those working without them, despite predicting beforehand that AI would make them 24% faster. The fundamentals of good software delivery have never been more important — and never more neglected. When instruments go dark The Navy’s answer to training dogfighters for success was the Top Gun school — not just to teach pilots to fight, but to teach them how to fly again. That meant returning to the fundamentals by mastering the technical and combat skill…
Truckloads of Tesla Batteries Keep Getting Stolen Before They Even Leave the Factory
Nine major suspected cargo thefts happened at Tesla’s Nevada battery factory in January alone, according to sheriff’s records obtained by WIRED.
The rise of the product engineer: How AI is reshaping modern tech teams
The end of pure specialization For years, software organizations optimized around specialization. Product managers owned requirements. Engineers owned implementation. Designers owned UX. QA owned quality. The model worked – until product velocity became a competitive advantage measured in weeks instead of quarters. Today, AI is accelerating another shift that I believe will fundamentally reshape how high-performing technology teams operate: the rise of the product engineer. As Chief Technology Officer of akirolabs, an AI-augmented strategic procurement platform serving enterprise-scale clients, including Fortune 500 organizations, I’ve spent the last several years evolving our engineering model through three distinct stages. First, I dismantled highly specialized silos. Then I transitioned the organization toward more flexible generalists. Eventually, our operating model revealed that the teams performing best in the AI era were neither traditional specialists nor pure generalists, but engineers deeply embedded in product thinking and business context. I formalized and operationalized this role internally as a product engineer model, adapting an increasingly common industry pattern to enterprise AI delivery. This role does not replace product managers. Instead, this operating model elevates strong product managers by removing operational friction. In our organization, product managers became more focused on customers, roadmap prioritization, requirement validation and strate…
The great cloud rebalance
For years, the enterprise narrative focused on moving to the public cloud for flexibility and leaving behind old infrastructure. While the public cloud remains a powerful platform for burst capacity, global reach, and modern application development, leaders now evaluate where each workload can achieve the best financial performance, operational efficiency, and risk. Cloud repatriation is back on the CIO’s agenda. Cloud repatriation does not always mean dragging workloads back into a company-owned data center. In many cases, enterprises are moving applications and data from hyperscale public cloud platforms into colocation environments, hosted private clouds , or MSP-operated infrastructure. The common thread is not nostalgia for on-premises IT. It is the desire for a more suitable workload placement. Enterprises are deciding that some systems belong in public cloud while others are better served in environments with more predictable economics, tighter control, and fewer architectural compromises. Cost is the loudest signal The most common reason enterprises repatriate workloads is cost. Public cloud pricing works extremely well when demand is variable, when teams need rapid provisioning, or when a business wants to avoid upfront capital spending. But not every enterprise workload behaves that way. Many core systems are steady, always-on, data-intensive, and relatively predictable. For those workloads, usage-based pricing can become less attractive over time. Compute charges,…
Europe’s industrial wake-up call
Europe’s climate ambitions must integrate with industrial competitiveness. The upcoming EU ETS reform should support decarbonization while ensuring affordable energy, investment certainty, energy security and a level playing field.
AI data infrastructure startup Clairva raises $500K led by Venture Catalysts
AI data infrastructure startup Clairva has raised $500K in a pre-seed funding round led by Venture Catalysts through its angel network. The company will use the fresh capital to strengthen its licensed data supply network, expand partnerships with content owners and institutions, enhance data enrichment and validation capabilities, and support commercial engagement with global AI customers, Clairva said in a press release. Founded in 2025 by Sunil Nair, Sabari Raju, Dushyant Verma, and Amit Parashar, Clairva builds licensed, provenance backed datasets for AI foundation models, embodied AI, robotics, and autonomous systems. As AI models increasingly rely on high quality datasets, sourcing data with clear usage rights, provenance, and cultural context remains a challenge. Clairva works with content owners, production houses, studios, archives, institutions, and contributor networks to source, license, and structure real world data for AI training. The company is initially focused on India, Southeast Asia, and other Global South markets, where languages, environments, behaviours, gestures, workflows, and objects remain underrepresented in AI training datasets. According to Clairva, it is also developing proprietary technology across the data pipeline, including licensed dataset ingestion, rights and provenance tracking, automated enrichment, metadata generation, action and object tagging, temporal segmentation, quality validation, and dataset packaging.
Prosus India posts $781 Mn revenue in FY26; turns adjusted EBITDA positive
Prosus reported a strong performance from its India business in FY26, with the ecosystem turning adjusted EBITDA profitable, driven by fintech platform PayU. According to the Prosus’ FY26 annual report, revenue from its India ecosystem rose 13% year-on-year to $781 million in FY26 from $694 million a year earlier. The business posted an adjusted EBITDA (aEBITDA) profit of $18 million, compared to a loss of $25 million in FY25, while its adjusted EBIT loss narrowed sharply to $10 million from $49 million. The Prosus India ecosystem employed 3,897 people during the fiscal. PayU's payments business generated $577 million in revenue,which grew 10% year-on-year, while delivering EBITDA of $12 million. Higher-margin value-added services (VAS) and SaaS offerings contributed 33% of payments revenue, supporting margin expansion. The company's credit business also reached profitability, reporting an adjusted EBITDA of $6 million. Credit revenue rose 19% year-on-year to $204 million, while new loan issuances touched $221 million during the year. During FY26, PayU processed total payment volume (TPV) worth $90 billion across its platform. The company said PayU accounts for around 25% of India's online payments industry revenue and manages $682 million in assets under management through its lending business. During the year, it also increased its stake in banking payments technology firm Mindgate to 70.7%. Prosus said it is building one of India's most comprehensive digital consumer ecos…
Gambia Trains Journalists on AI in Media as African Union of Broadcasting Assembly Convenes in Banjul
The Gambia’s Ministry of Information, Media and Broadcasting Services, in partnership with the African Union of Broadcasting, has launched a two-day capacity-building programme for Gambian journalists on artificial intelligence and broadcasting, held as part of the 17th General Assembly for journalists in Banjul from April 13 to 16. The training, under the theme “Artificial Intelligence [...]
Zambia Police Warn Against AI-Generated Misinformation Targeting Officials Ahead of 2026 Elections
The Zambia Police Service has warned the public against the creation and circulation of misleading AI-generated content depicting violence and other offensive material aimed at senior public officials. Police spokesperson Godfrey Chilabi said the service had noted with concern the misuse of artificial intelligence to produce deceptive digital content. In a statement issued in Lusaka, [...]
Heatwave: Berlin police use water cannon to cool Bruno Mars fans
Berlin police used a water cannon to cool thousands of Bruno Mars fans queuing outside the Olympiastadion as temperatures soared during Europe's heatwave.
Snupit Leverages Locally Built AI to Connect South Africans With Trusted Service Providers
Artificial intelligence continues to reshape industries worldwide, and South Africa’s online services sector is no exception, with AI increasingly used to help customers find the right businesses faster and more efficiently. Online services marketplace Snupit has been leveraging AI-powered technology for years to improve how customers connect with trusted local service professionals. While many companies [...]
Concern over AI’s impact on Asia’s real estate sector is misplaced
Artificial intelligence (AI) is a transformative force in economies and financial markets. McKinsey estimates that data centres alone will need a staggering US$6.7 trillion in capital investment across the global data centre value chain in the next five years, and the scale of the investment suggests the world economy is in the early stages of a far-reaching shift that could lead to big gains in productivity. In a report on June 24, Vanguard said the massive buildout of AI infrastructure...
Telling internet platforms where to stick public service media will serve nobody. Turn it on its head
Show not tell makes a far finer script
Samsung and SK Hynix plan $590 billion chip investment as AI demand sends memory prices soaring
Samsung and SK Hynix, backed by the South Korean government, are pouring $590 billion into new chip factories and packaging centers as AI data center demand surges. According to Jefferies, memory prices could climb to 50 percent per quarter through 2027. The two companies control nearly 80 percent of the global HBM market. The article Samsung and SK Hynix plan $590 billion chip investment as AI demand sends memory prices soaring appeared first on The Decoder .
Watch: Lightning hits Eiffel Tower during violent storm in Paris
Lightning struck the Eiffel Tower as a powerful thunderstorm swept across Paris on Saturday night, sending visitors scrambling for shelter. The dramatic weather followed days of record-breaking heat in France, where authorities say the extreme temperatures contributed to a sharp rise in deaths.
Proper way to split dataset (split ratio) and evaluate the baseline model or fitted model in the firsthand
I appreciate your time in this matter. I really need an answer for this, for my thesis. I use the PaySim dataset from Kaggle ( https://www.kaggle.com/datasets/ealaxi/paysim1 ). First of all, I use training, validation, and testing set. Is there really a rule on data split ratio and is it acceptable if I check the model performance on each split, for example, 70/15/15, 80/10/10? After fitting the model with training dataset, we get the default model/fitted model. Which dataset (Training or Validation set) shall I use for examining the model performance? My intention of having 3 types of set (Training, Validation, and Testing) is to use the Validation set as hyperparameter tuning examination. Thanks so much.
Why enterprises aren’t ready for AI Agents yet
Enterprise AI isn't scaling as fast as many expected. Here's what's really holding businesses back, and why it matters more than ever. The post Why enterprises aren’t ready for AI Agents yet appeared first on MEDIANAMA .
Introducing GPT-5.6 series: Sol, Terra and Luna
So we will probably get it in about 3 months when open models will likely have surpassed it
Info session: Call for proposals – Youth-driven social media platforms
Info session: Call for proposals – Youth-driven social media platforms Anonymous (not verified) Mon, 06/29/2026 - 10:03 30 July 2026 Online, 10:00 – 12:00 CEST Join this info session dedicated to all interested potential applicants of the new call for proposals to shape safer and more inclusive social media platforms features. © Worawee Meepian The project will involve young European people in the design of features for new social media platforms, focusing in particular on user well-being, mental health, safety, privacy, and ease of use. No registration is needed. Find out more about the call for proposals . Join this meeting online Contact mailto:CNECT-I4@ec.europa.eu Related topics Media and democracy Better Internet for Children Strengthening trust and security Online platforms and e-commerce Social media and networks, innovation and policy Funding for Digital
DiScoFormer: One transformer for density and score, across distributions (duplicate)
DiScoFormer is a transformer-based density and score estimator that can infer both quantities from a finite sample in one forward pass, generalizing classical KDE while staying accurate in high-dimensional and out-of-distribution settings without retraining for each new distribution.
European tech weekly recap: €2.1B in deals and Tech.eu Funding Explorer
Last week, we tracked more than 75 tech funding deals worth over €2.1 million and over 5 exits, M&A transactions, rumours, and related news stories across Europe. 📊 The top three industries that r...
This Humanoid Robot Is a Terrifyingly Competent Office Intern
Flexion Robotics, a startup founded by ex-Nvidia engineers, has a clever way of training robots to do useful work.
Renfe cancels 320 trains on Monday due to rail strike in Spain: thousands of passengers hit
The 24-hour strike is disrupting high-speed, regional and suburban rail services just as the holiday getaway is getting under way.
As AI’s Novelty Fades, Cannes Lions’ Winners Point to What’s Next
This year’s Cannes Lions was an inflection point where ‘AI stopped being the headline and simply became part of the creative process.’
US and Iran agree to resume talks after weekend of strikes
Washington has proposed this week's discussions, which are expected to center on the Strait of Hormuz, take place in Doha and begin as early as Tuesday.
One killed in shooting at California World Cup fan zone, San Jose police say
The shooting happened at San Pedro Square, one of several places in the San Francisco Bay Area where crowds have gathered to watch World Cup matches on big outdoor screens.
Telegram Most-Used App For Drug Advertising, NCB Report Says
NCB has identified Telegram as the most prominent encrypted platform used to advertise illegal drugs in India. The post Telegram Most-Used App For Drug Advertising, NCB Report Says appeared first on MEDIANAMA .
Boundless Ventures leads pre-seed round in Lytmus AI
Lytmus AI, a Bengaluru-based startup building AI mentors for competitive exam preparation, has raised Rs 5 crore in a pre-seed funding round led by Boundless Ventures. The fresh capital will be used to strengthen the startup’s AI capabilities, accelerate product development, and expand student acquisition, with an initial focus on the NEET segment, the company said in a press release. Founded in 2024 by Ajit Kumar and Praveen, Lytmus AI develops AI mentors for competitive exam preparation. The platform combines persistent memory and contextual understanding with subject expertise from teachers to create AI mentors that resolve doubts in real time and guide students through their learning journey. It is currently available on mobile with an initial focus on NEET. According to the startup, its platform operates on two layers. The infrastructure layer is trained on the teaching patterns of subject experts to enable AI mentors to explain concepts effectively, while the application layer maintains memory and context for each student to personalise the learning experience. With more than two million candidates registering for NEET every year, the company believes there is significant demand for personalised academic support during exam preparation. Lytmus AI said its platform has been used by over 16,000 students in the past 90 days. According to the company, students using its AI mentor complete up to three times more daily practice than those without such guidance. The startup c…
Heatwave: Which European countries are running out of water?
As Europe endures another brutal summer, Europe in Motion unpacks water stress levels across the continent.
앤트로픽 출신 베테랑들, ‘과학자용 AI’ 스타트업 설립…역대급 시드 투자 유치
앤트로픽 출신의 핵심 연구원들이 설립한 AI 스타트업 미렌딜(Mirendil)이 대규모 투자 유치에 성공하며 본격적인 행보에 나섰다. 오픈소스 개발자와 과학자들이 자체적인 AI 모델을 직접 구축할 수 있도록 돕는 \'AI를 만드는 AI\' 개발이 목표다.월스트리트저널(WSJ)에 따르면, 미렌딜은 최근 앤드리슨 호로비츠와 클라이너 퍼킨스, 엔비디아 등으로부터 2억달러(약 3083억원) 규모의 시드 투자를 유치했다. 이번 투자에서 미렌딜은 기업 가치 10억달러(약 1조5414억원)를 인정받으며 단숨에 유니콘 기업 반열에 올랐다. 최근 신생
칼럼 | 미토스가 던진 질문, CISO들이 이사회에 꺼내야 할 진짜 답
새로운 AI 기술이 세간의 관심을 끌 때마다 이사회와 경영진 회의에서는 늘 비슷한 장면이 반복된다. 새로운 기술은 곧 공격자에게 강력한 무기가 될 혁신으로 소개되고, 논의는 빠르게 최악의 시나리오로 흘러간다. 이어 보안 책임자들은 어김없이 같은 질문을 받는다. ‘이전에는 없던 새로운 위험에 갑자기 노출된 것인가?’ 필자의 대답은 대체로 ‘아니다’이다. 대부분의 조직에서 더 큰 문제는 미토스 같은 최첨단 AI 모델이 하루아침에 새로운 위험 범주를 만들어낸다는 점이 아니다. 오히려 이러한 모델은 사이버 보안의 공격과 방어 모두에서 업무 속도를 높일 수 있다. 공격자는 더 빠르게 공격을 수행할 수 있고, 방어자는 오랫동안 방치돼 온 취약점을 더 신속하게 찾아 우선순위를 정하고 해결할 수 있다. 그래서 필자는 미토스를 사이렌(siren)이 아니라 신호(signal)로 본다. 이는 사이버 공격과 방어의 경제성이 변화하고 있음을 알리는 신호일 뿐, 보안의 기본 원칙이 더 이상 중요하지 않다는 의미는 아니다. 오히려 그 반대다. 자산을 명확하게 파악하고, 체계적인 패치 관리와 강력한 ID 통제, 복원력 있는 운영 체계를 갖춘 조직일수록 앞으로 AI가 어떤 변화를 가져오더라도 훨씬 유연하게 대응할 수 있다. 이러한 관점이 중요한 이유는 최근 발생한 침해 사고 역시 익숙한 실패 지점을 보여주기 때문이다. 버라이즌의 ‘ 2025 데이터 침해 조사 보고서(Data Breach Investigations Report )’에 따르면 자격 증명 탈취와 취약점 악용은 여전히 조직이 침해되는 주요 원인으로 꼽히며, 특히 취약점 악용 사례는 계속 증가하고 있다. 다시 말해 기업 내부로 침투하는 경로는 대부분 보안팀이 이미 잘 알고 있는 취약점에서 시작된다. 진짜 문제는 여전히 ‘기본’ 필자의 경험상 많은 조직이 겪는 문제는 전략의 부재가 아니라 실행력 부족이다. 보안 리더는 기본 원칙을 알고 있고, 실무팀도 알고 있으며, 감사기관과 규제기관, 이사회 역시 이를 잘 알고 있다. 문제는 하이브리드 환경, 노후 시스템, 클라우드 플랫폼, 원격 근무 환경, 복잡한 서드파티 생태계 전반에서 이러한 기본 원칙을 일관되게 유지하는 일이다. 이 때문에 AI가 보안 통제의 우선순위를 근본적으로 바꿀 것이라는 전망에는 신중한 입장이다. 대부분의 성공적인 침해 사고는 여전히 이미 알려진 취약점에서 시작된다. 단지 적절히 조치되지 않았거나, 우선순위가 잘못 설정됐거나, 애초에 존재조차 파악하지 못했던 취약점일 뿐이다. 인터넷에 노출된 미패치 시스템, 잘못 구성된 ID 관리 체계, 과도한 권한, 미흡한 네트워크 분리, 수년간 검토되지 않은 서비스 계정, 그리고 일시적 예외로 시작했지만 어느새 영구화된 핵심 업무 예외 등이 대표적이다. 필자는 보안 프로그램이 최신 위협에 지나치게 집중하면서 추진력을 잃는 사례를 여러 차례 봤다. 기존 통제 공백은 그대로 둔 채 새로운 활용 사례에만 예산을 투입하고, 책임 체계와 운영 프로세스, 관리 체계를 바로잡기보다 새로운 보안 도구부터 도입한다. 또한…
[게시판] 한컴, 한국서부발전에 한컴어시스턴트 공급 등 단신
■ 한컴(대표 김연수)은 한국서부발전의 생성형 AI 챗봇 \'위피봇\'에 AI 문서작성 솔루션 \'한컴어시스턴트\'를 공급한다고 밝혔다. 한컴어시스턴트는 한국서부발전이 보유한 사규·법령·업무 매뉴얼·안전자료 등 약 72만건의 내부 지식 데이터와 연계된다. 이를 통해 임직원은 보고서 작성 과정에서 관련 규정과 업무 정보를 실시간으로 활용해 정확하고 일관된 문서를 작성할 수 있다는 설명이다.■ 셀바스AI(대표 곽민철)는 약 5만명의 의사회원을 보유한 의료인 전용 지식·정보 공유 플랫폼 \'인터엠디\'와 협력해 AI 음성인식 솔루션 \'셀바스 메디보
[게시판] 위로보틱스, 휴머노이드 ‘알렉스’ 시뮬레이션 모델 공개 등 단신
■ 위로보틱스(대표 이연백, 김용재)는 피지컬 AI 개발 생태계 구축을 위한 기술 공개 로드맵을 시작한다고 밝혔다. 첫 번째 공개 기술로는 휴머노이드 ‘알렉스(ALLEX)’의 시뮬레이션 모델과 심-투-리얼(Sim-to-Real) 검증 결과를 공개했다. 이번 모델을 통해, 개발자와 연구자는 실제 하드웨어 없이도 알렉스 하드웨어 기반의 제어, 학습, 합성 데이터 생성 연구를 수행할 수 있다.■ 아이스크림에듀(대표 이재준)는 28일 수원시 청소년문화센터 온누리 아트홀에서 열린 ‘수원새빛인강 학부모 설명회 및 오리엔테이션’에 참여해 AI
Power shift: Will Burnham’s big plan work
Barring any last-minute surprises, Andy Burnham will become prime minister three weeks today. But what is his 10-year mission for the country and how will he achieve it? In his first major speech in Manchester, Burnham will set out his plans for the economy, welfare and devolution with a pledge to deliver “good growth in […]
Why searching old ChatGPT conversations becomes impossible at scale
If you have been using ChatGPT as a daily Work tool for more than a year, You’ve probably run into this: finding a specific old conversation is not actually possible through search. You scroll. You scroll more. At some point, you reconstruct the work from scratch instead. This is not a minor friction point. Add a certain volume of conversations - somewhere around 200 to 500, depending on how heavily you use the platform - the sidebar stops functioning as a navigation tool and becomes an archive You cannot access in any practical sense. ** What the interface give you** Chat GPT organises conversations into time buckets: today, yesterday, previous 7 days, previous 30 days, and then individual months going back. There is no search bar that searches conversation content. There is no way to filter by topic, keyword, or project. The only retrieval method is remembering roughly when a conversation happened, scrolling to that bucket, and scanning title. Titles are generated automatically. They often do not reflect what the conversation contained. A session where you worked through a complex data problem might be labelled “ python data analysis” - the same label that could apply to a dozen other sessions. At scale, these labels stop being useful identifiers. ** what costs in practice** The most common failure modes: Prompt you developed over several exchanges - one that worked well for a specific task - cannot be located when you want to reuse it. A client asks about the reasoning be…
YOLO11-seg underperforming EfficientNet-UNet for building footprint extraction from aerial imagery – what should I try next?
I'm looking for advice from people with experience in remote sensing and instance/semantic segmentation. I'm working on building footprint extraction from aerial imagery. I have a baseline segmentation model based on EfficientNet-B7 U-Net, which performs reasonably well on my test areas. I wanted to explore whether a YOLO segmentation approach could provide competitive results, so I fine-tuned a YOLO11 segmentation model. The results, however, are significantly worse than my U-Net baseline, and I'm trying to understand whether this is expected, whether I'm using the model incorrectly, or what I should try next. Dataset Task: single-class building footprint extraction Imagery: high-resolution aerial/satellite imagery (~50 cm GSD) Training images: 891 for fine tuning, I have used 12k for pre training the model) Validation images: 156 (for fine tuning, I have used 2155 for pre training the model) The model was initialized from weights previously trained on a large building footprint dataset and then fine-tuned on my local dataset. Training configuration Model: YOLO11m-seg Epochs: 100 Best epoch: 78 Image size: 640 Batch size: 16 Initial LR: 0.0005 Cosine scheduler: enabled Mosaic: 0.5 Rotation augmentation: ±90° Horizontal flip: disabled Vertical flip: disabled Patience: 20 Best validation metrics Box metrics: mAP50 = 0.6438 mAP50-95 = 0.3894 Mask metrics: mAP50 = 0.6345 mAP50-95 = 0.3236 Precision(M) = 0.7436 Recall(M) = 0.5957 Inference observations One thing that concerns me…
