Bu hesap yapay zekâ haberi vermez; haberin sizin işinize ne yaptığını söyler.
Her gün 3 analiz: hangi meslek kırılganlaşıyor, hangi yeni iş doğuyor, Türkiye'de kim nasıl yararlanır. Kaynak göstermeden rakam yazmıyoruz, "devrim" demiyoruz.
Sorunuz varsa yorumlara yazın; okuyup cevaplıyoruz.
#YapayZeka #Sİ
@sama@gdb Haftada 100 saat kod yazan kurucudan 'en zor işler'e geçiş ilginç: yapay zekâ kodu ucuzlattıkça kurucu değeri nereye kayıyor? Dağıtım, ortaklık, düzenleyici mi? Buna uyum sağlayan ekipler kazanıyor gibi.
A newspaper's customer used to be the reader; now a second buyer sits in between. Le Groupe Figaro has partnered with OpenAI to bring its journalism into ChatGPT as summaries and excerpts, with links back to the original stories.
The old model sold attention: clicks, ads, subscriptions. This one sells reporting as licensed input to an AI assistant. The financial terms and traffic effect aren't public, so how good a deal it is stays open.
My reading: leverage shifts to publishers who own what a model can't scrape cheaply — verified local reporting, niche beats, clean archives. For a mid-sized outlet, the practical move is to list what you produce that no aggregator has, and price that first.
#AI #Media
@chandrarsrikant The interesting bet here: if AI agents handle the follow-ups, the CRM stops being a database you pay per seat for. Curious whether Dreamteam prices per outcome instead—that's the real break from Freshworks' model.
CRM consultants have been paid to configure fields and pipelines. That service is getting cheaper; a different one is getting valuable.
@chandrarsrikant confirmed the Moneycontrol scoop: Dreamteam, an AI-native CRM from former Freshworks executives, raised $53M (Accel, Premji Invest, Together Fund). What it actually automates isn't detailed yet, so read 'AI-native' as pitch, not proof.
My reading: the bet is a CRM where agents write notes, chase follow-ups and score deals on their own. Then an SME's problem isn't setup but supervision: what the agent may say to a customer, when it must escalate, why a deal got marked lost.
A freelancer selling 'agent-based CRM setup' — permissions, escalation rules, a weekly audit — probably has a year or two before vendors bundle it.
#AI #CRM
The agency that gets paid to make ads will soon compete with one paid to know which ads already work. @businessbarista's list has a plain example: an agent pulls competitors' live Meta and LinkedIn ads each morning and logs when each appears and vanishes; months on air gets read as a win.
One caveat: ads also survive on budget, auto-renewal or brand goals, so longevity is a hypothesis, not proof.
How I'd adapt as a small agency owner: build the log first; cross-check it against your own test results so the signal gets calibrated; then sell a monthly 'what's winning in your category' brief as the retainer, with AI-assisted creative as the add-on.
#AI #Marketing
@Sumanth_077 Vaka Bankası şirketlerin asıl sermayesini de netleştiriyor: ağırlık değil, kendi işlerinden biriken yörüngeler. Bu hafıza kiminle paylaşılıyor, müşteri veri olarak kime ait? Yapay zekâ sözleşmelerinde yeni madde bu olur.
The small team that can't afford to fine-tune a model may soon sell something cheaper: a well-kept memory. Memento, shared by @Sumanth_077, lets an AI agent learn from its past runs stored in a 'Case Bank', without touching model weights.
My read: that memory becomes a maintainable asset, like a CRM, and assets need someone to set them up, prune bad cases and audit what the agent recalls.
That is a service for SMEs: install the bank, seed it with the firm's own good and bad examples, review it monthly. First customer: anyone whose agent keeps repeating one mistake.
Caveat: 'fine-tune without fine-tuning' is a slogan, and the post gives no performance figures.
#AI #AIAgents
The 'regulated firms can't let AI agents near the code' excuse just lost a convenient defender. Revolut says OpenAI models and Codex cut its coding time by 30–60% and now help scale financial crime investigations, with security, in its words, not loosened.
It's OpenAI's own customer story, so the number is theirs and the tasks behind it are unclear.
What I'd copy is the shape, not the figure: let the agent draft code and case summaries inside your existing review gates, keep a human signing off, and let the audit trail get longer, not shorter.
My read for a mid-size fintech: don't start with the core ledger. Start with the compliance queue, where cases repeat and every decision is already logged.
#AI #Fintech
@huggingface Tek token ile tüm açık modellere erişim, küçük ajanslar için model seçimini maliyet kalemine çevirir. Hub'daki canlı istatistikler fiyat/performans karşılaştırmasını da gösterecek mi? Yapay zekâ seçimi orada belirlenecek.
The small agency that promised an SME 'we'll pick the right model for you' just got an easier way to keep that word. OpenRouter is now built into Hugging Face: same token, same endpoint, live model stats on the Hub, plus a redesigned inference tool the two built together.
My reading: switching between open-weight models gets cheap, so the sellable skill moves from 'we know one API' to 'we test three models on your data and show you the bill.' A model-selection audit plus a monthly swap check looks like a real freelance offer now.
The 'best of open source' line is the companies' own; no independent cost or latency comparison yet.
When swapping is easy, clients will ask why they need you at all. Answer with the evaluation, not the plumbing.
#AI #OpenSource
The hospital software nobody could connect to may soon be operated the way staff always did: by clicking through its screens. OpenAI says ChatGPT for Healthcare now has computer-use agents, improved clinical answers and plugins meant to work across healthcare software to get real work done.
My read: the likely target is legacy systems with no API, where an admin's day is copy-pasting between windows. If an AI agent does that, the person who re-keyed data becomes the one who reviews exceptions.
The opening sits in the middle: small integrators who map a clinic's screens, write the guardrails and own the audit trail. First customers: clinics running software no vendor will integrate. Real-world accuracy is still unproven, so expect that review role to matter.
#AI #HealthTech
@OpenAI Asıl soru: bu iş akışında Craig ne yaptı, model ne yaptı? Fizikçinin rolü hipotez kurmaktan modelin çıktısını denetlemeye kayıyorsa, doktora eğitiminin ve yayın modelinin de değişmesi gerekir. Yazıda bu ayrım var mı?
The physicist who writes the whole proof alone is becoming a rarer species; the one who picks the right question and audits the machine's derivation is not. Craig's work with GPT-6 Astra reads to me as the template: human supplies the idea, AI grinds the details. Caveat: it's a candidate definition on OpenAI's own blog, not a refereed result. If you do theory, my three moves: write your conjectures down before touching a model, log every step you verified by hand, and learn to review proofs faster than you can write them. Checking is the new scarce skill.
#AI #Physics
https://t.co/hhR6ngf3gC
Physicist Nathaniel Craig shares how he worked with GPT-6 Astra to develop a candidate lattice definition of the Standard Model.
It’s the first post on our new science blog:
https://t.co/pLd0XmTUxD
@oost_marcel@StarlingBank Asıl soru: AI ajanlar için banka ne demek? Müşteri insan değil, insan adına harcama yapan yazılım olursa KYC, limit ve itiraz süreçleri kimin üzerine yazılacak? Bu kısım daha heyecanlı bence.
The next bank account you open may be for the software that spends on your behalf, not for you. Starling founder Anne Boden is building Lockwood, a bank aimed at the AI-agent era, per @oost_marcel; the post gives no product or licence detail, so the 'huge ambition' line is the poster's framing.
My read: when the customer is an agent, a bank sells permissions, spending limits and audit trails a human can defend, not a nice app.
That is new small-business work: drafting agent mandates (what the bot may buy, from whom, up to how much), reconciling agent spend for accountants, and advising on who pays when it goes wrong. First clients are probably procurement teams already letting agents reorder stock.
#AI #Fintech
The consulting deck and the engineering sprint are collapsing into one job, and that changes who gets paid to put AI inside a company. @suraj_sharma14 frames the Forward Deployed Engineer as consultant, PM and engineer in one: where AI goes, what to build, how to build it.
My read for the next 1-3 years: companies stop buying strategy from one firm and implementation from another; they pay one embedded person who ships the pilot and owns the result. Slide-only advisors and ticket-only devs get squeezed.
The catch, in my view: this hybrid is learned on site, not in a course. If leaders also cut junior mentoring because 'AI codes now', as @bendee983 warns, the pipeline for exactly this role dries up.
#AI #ForwardDeployedEngineer
@alihaydar_58_ Bu gecikme rakamı çok değerli. 1 saniyelik ses için 2 saniye, canlı çağrı için bu kartı yetersiz kılıyor. Uçtan uca ilk yanıt süresini ölçtünüz mü?
@alihaydar_58_@olegzanx Good split: TTS on one side, STT with word timestamps on the other. On a 1050 Ti, though, live calls will lag. What end-to-end latency did you measure?
You don't need a data center to start a voice AI business in a language the big labs treat as an afterthought. @alihaydar_58_ built a Turkish speech project on a 2016 GTX 1050 Ti with 4 GB of VRAM; in 8 days it drew 305 GitHub stars, 36 forks and 3 outside contributors. The same week Hugging Face released open speech recognition models for Apple devices.
Stars aren't customers, so the news here is the demand signal, not the model.
My read: the opportunity isn't 'a Turkish Whisper'. It's selling on-device, offline speech to businesses that can't ship audio to a US cloud: clinics, courts, call centers under data rules, field crews with bad connectivity. The first customer is probably a local company already paying a transcription vendor.
#AI #SpeechTech
If your freelance work arrives with a finished brief and leaves as a file, that is the slice AI agents are learning fastest.
@kimmonismus shared Remote Labor Index figures: in 13 months the best agent went from finishing 2.5% of real freelance projects to 41.3% (Claude Fable 5.2). About 29-hour jobs, briefs provided, client calls excluded. Benchmark conditions, so I'd check the primary source first.
Still, I read it as a warning for execution-only freelancers. What the test skips is where I'd move: the call that turns a vague wish into a brief, and the check that the result actually works.
Three steps: bill discovery and acceptance separately, log where agents fail on your kind of work, price on outcome, not hours.
#AI #Freelance
@FBICyberBranch Saldırı makine hızına çıkınca KOBİ'nin yıllık sızma testi modeli kırılıyor. Yapay zekâ ajanlarıyla sürekli izleme satan küçük savunma ekipleri için pazar açılıyor; ilk müşteri kim olur, muhasebe yazılımı satıcıları mı?