This compilation of December 2024 clips runs from practical advice on building with LLMs to a sceptical frame for AI itself. It starts with build first and optimise later, moves through model selection and data security in agentic systems, and ends with a guest arguing that the industry's certainty about AGI is a faith claim, and that slowness may be a virtue.
Cost and model choice. Andrew Ng (episode 841) says teams worry too much about API prices. A weekend day of his own coding ends at about a $5 OpenAI bill, and agentic workloads run to tens or low hundreds of dollars. His advice is to use the best model, build something valuable, and only then optimise with smaller models or supervised fine-tuning. Ed Donner (episode 847) adds that an AI engineer's first job is choosing the right LLM from data, business-outcome metrics and budget and time constraints. He recommends starting from a non-LLM baseline, as he and the host did with heuristics and logistic regression at their earlier startup, and then usually a closed-source model. He lists four reasons to go open source: proprietary data, sensitive data, inference cost, and on-device use.
Security in agentic systems. The episode 843 guest from Protopia argues that agents spanning on-prem, private and public clouds force component-by-component thinking about data exposure. He cites homomorphic encryption work, including a recent Apple announcement, and partnerships across the stack.
2024's highs and lows. In episode 849 the host and his co-guest found OpenAI both the year's disappointment (hers) and its wow moment, o1 (his). She calls this "TikTok-ified" expectations. Her wow was NotebookLM; his near-wow was Waymo. He forecasts disruption of truck driving, the top job in most US states, but expects automation to raise employment as it has before.
Tech agnosticism. Greg Epstein (episode 845) recasts the host's AGI optimism as a slow multi-stage Rapture. He argues that confident figures sell certainty in an uncertain world, that Sam Altman's $7 trillion ask reflects OpenAI's cash burn, and that losing the arms race might be worth it.
Coleman Hynie, CEO of Counterpoint, ends on communication: nearly anything, she says 99.9% of things, can be worked out by talking honestly and acknowledging people and privilege, and work would be less lonely if people spoke less formally. Her product is pitched at relieving the tension between how managers perceive value and what employees contribute.
A boss who cares. Hanselman opens with his manager rejecting his Microsoft review as unclear; he rewrote it more quantitatively and credits a decade with a caring boss, which he says almost no one has.
What happened to managers. She describes engineers pushed into management who compete with their own reports, a 50/50 ideal split between contribution and people work that no longer exists, and reporting that absorbs the role. Hanselman cites 40-page documents read together in meetings.
Numbers and stories. Metrics like bug counts get used to tell the story leadership wants. When top-down communication drops context, employees stop owning outcomes; she asks whether companies should "loosen the strings" and share real business needs.
Generations and job-hopping. Hanselman admits a bias against 18-month stints; she says people do what high rents require, but that the pattern yields managers without people skills, and that the manager title carries outsized family and cultural prestige.
Adults and accountability. She backs treating staff as adults who can take critique and be held accountable, contrasting her upbringing's directness with Hanselman's conflict-averse Gen X style. Feelings belong at work, she says, alongside responsibility.
Identity. She describes code-switching and losing her identity climbing as a Black woman, and the privilege senior executives have to be quirky.
Karan Singh, an engineering leader and senior software engineer at Salesforce, walks the Data Neighbor podcast hosts through how AI reached large language models and agents. The position Karan ends on: today's LLMs are powerful retrievers that cannot yet reason, agents and RAG work around that, and artificial general intelligence follows once reasoning is solved, which Karan predicts within 5 to 10 years.
A career built on pivots and contacts. Karan moved from Android development, through a master's in software engineering with a data minor, into deploying machine learning models under compliance constraints at a fintech, and then to bringing AI into Salesforce's infrastructure layer. Both later moves came from phoning former managers, which the hosts present as the case for keeping early connections.
From rules to Transformers. Karan traces 1960s rule-based systems, probabilistic models, traditional machine learning such as linear and logistic regression, and neural networks that infrastructure held back until deep learning. A neuron is described as a linear unit with parameters, like m and c in y = mx + c, passed through an activation. Transformers allowed parallel training, which is why only the largest companies with infrastructure build frontier models.
Agents and jobs. An agent gives the LLM "brain" tools: a writer, an email sender and an SMS sender replace a month of if-else code. Karan argues that software engineers are still needed, with humans in the loop, and compares the shift to moving from PHP to React. The hosts add that creativity and domain knowledge decide what anyone can get out of these tools.
Retrieval is not reasoning. An LLM resembles a reader with photographic memory: trained on bad data, it answers anyway, and hallucination is sequence completion. RAG grounds answers in supplied knowledge and allows source tracing. Karan says AGI arrives once reasoning works, citing ReAct-style prompting and OpenAI's new reasoning models, alongside responsible-AI work. For learners the episode recommends DeepLearning.AI, Andrej Karpathy's videos, agent frameworks, and a daily AI news article for the non-technical.
Florian Neukart of Terra Quantum argues that quantum machine learning is practical now, provided it is hybrid. Classical computers do what they are good at, quantum circuits are added only where they help, and the result is judged against the best classical solver. He ends on a security stack he calls "absolutely secure" against quantum attack, a room-temperature superconductivity result from Terra's own team, and an invitation to anyone with a machine learning background to start without a physics degree.
Hybrid rather than all-quantum. The host cites episode 721, where an earlier guest doubted near-term quantum ML. Neukart disagrees: early work shrank real problems into toy ones. Terra instead expresses parts of a network, such as a fully connected layer, as quantum circuits, stacked gates run on a gate-model machine. Its software maps those gates to whichever chip a customer picks, so switching hardware is a parameter change. Circuits must be shallow and touch few qubits at once, because coherence lasts nanoseconds to microseconds. Each added qubit doubles the configurations held in superposition. Because results are probabilistic, a circuit is measured perhaps a thousand times and the majority answer taken.
Applications. He groups uses into optimization, machine learning and simulation. Scheduling is the "boring" one he calls intractable classically: for Volkswagen, fitting 250-odd test classes to staff skills on cars leaving the line, and the same algorithm can schedule EV software subroutines to save power. Other cases are logistics (perishable goods at hubs within time windows, fewer empty miles), satellite constellations, collateral optimization and exotic options pricing. Simulation of molecules for batteries and drugs is the hardest, and he says it awaits stronger systems.
Hardware and Willow. Chip choice usually follows existing cloud contracts with AWS, IBM or others; IBM, Rigetti, D-Wave and Google are all in play. Better chips mean more and better qubits, longer coherence and more connectivity. Asked about Google's Willow, Neukart says even very high two-qubit fidelity compounds into random output at millions of qubits, and Willow showed errors can be held down while scaling. Engineering and cooling challenges remain.
Getting enterprise-ready. Terra benchmarks against CPLEX or Gurobi and says that if it is not better it makes no sense to use. New problems sometimes need new algorithms, which limits scale for a young, small field. Many firms want on-premises computing, but superconducting systems cost 15 to 20 million and date quickly. Neukart argues that data embedded as zeros and ones is meaningless without the encoding, which eases cloud worries. Terra offers Nvidia's hybrid quantum cloud among its backends and relies on GPU tensor networks, as in its chemistry package for conformer search. These can later be translated into quantum circuits.
Encryption. The host quotes Satya Nadella's billion-years-versus-two-minutes claim for RSA-2048 and expects no "encryption apocalypse". Neukart lays out post-quantum cryptography (NIST's recent three algorithms, safe "as of current understanding"), quantum key distribution over fibre, and quantum random number generation.
Ethics, superconductivity and the future. Terra is about to start self-funded clinical trials of laser treatment for osteoarthritis and arteriosclerosis, with parameters tuned by a hybrid quantum algorithm. It also published a fatty-liver classifier that beat the benchmark. Neukart says Terra's team around CTO Valery Vinokur achieved room-temperature superconductivity at ambient pressure in graphite. Designing qubits around it is still to come, and he cautions that earlier superconductivity discoveries did not deliver quickly. He is most excited by personalised drug design, then fusion and materials such as a space elevator.