5 AI Myths & The Truth Behind Them: ML, Context, Agents & More

Comments · 73

• MI

  @MikeWoot65 · 2 months ago

  Wow, what a great teacher. More of this guy!

  24

• TG

  @tgothe418 · 1 month ago

  As someone who is constructing a hobbyist level LLM/AI project these videos are absolutely transformational for my understanding of what can be done. Thank you so much to IBM and Martin for sharing this information!

  2

• 36

  @360captureit · 2 months ago

  Another quality session from Martin - Thank you

  5

• RO

  @Roof-Crane · 2 months ago

  You've got a great teaching style: friendly, helpful, "translates" jargon into ordinary words.  Thank you.

  1

• FI

  @FieldReckoner · 2 months ago

  The compounding-error math around <a href="https://www.youtube.com/watch?v=OWPRU_Pc4Ng&amp;t=732">12:12</a> is the most under-appreciated point in the video. 95% per step sounds fine until 20 chained steps lands you near 36%. But the takeaway isn&apos;t &quot;don&apos;t automate&quot; — it&apos;s that checkpoints should be placed by consequence, not spread evenly.

  3

• DE

  @DeepCatAI · 2 months ago

  Totally agree. The sycophancy point is the one most people miss. Every time I push back on a model it just agrees with me instead of getting better.

  1

• FL

  @Flapjackers · 2 months ago

  All Ai still hallucinates when I use it, all models, 3-5 prompts, and sometimes randomly on its own.

  11

• DP

  @dphandle_ · 2 months ago

  Great content as always! Thanks!

• JI

  @jitsathathatavakorn143 · 2 months ago

  Thank you 🙏<br>Love to learn from this channel.

• MA

  @MachielGroeneveld · 2 months ago

  Most myths stem from overextension of LLM capabilities. The biggest myth is that all AI limitations will be ‘solved’ within 2-3 years.

  4

• LX

  @lxndrlbr · 2 months ago

  Chapters<br><a href="https://www.youtube.com/watch?v=OWPRU_Pc4Ng&amp;t=72">1:12</a> AI models hallucinate <i>not that</i> frequently<br><a href="https://www.youtube.com/watch?v=OWPRU_Pc4Ng&amp;t=231">3:51</a> Chain of Thought is <i>not</i> actually faithful to the computations that lead to the output<br><a href="https://www.youtube.com/watch?v=OWPRU_Pc4Ng&amp;t=360">6:00</a> Training is <i>actually</i> only one third of total AI compute (cycles, watt, ?) by end 2026<br><a href="https://www.youtube.com/watch?v=OWPRU_Pc4Ng&amp;t=508">8:28</a> Large context window are <i>not really useful</i> to search large information pools for multiple data pieces (needles in proverbial haystack)<br><a href="https://www.youtube.com/watch?v=OWPRU_Pc4Ng&amp;t=642">10:42</a> AI agents <i>cannot</i> work fully autonomously, human supervision required (human in the loop, verifier checks)

  6

• AN

  @AndrejsBoka · 1 month ago

  Yep, thanks, especially the myth about hallucinations—it really misleads users. Right now, intentional fine-tuning that introduces biases is probably way more dangerous in AI. And the part about &apos;thinking&apos; AI is so true, because as I understand it, the model generates those questions internally just to get closer to the topic of the prompt. But that’s not the model’s actual thought process.